Telecom Churn Case Study Hackathon

In the telecom industry, customers are able to choose from multiple service providers and actively switch from one operator to another. In this highly competitive market, the telecommunications industry experiences an average of 15-25% annual churn rate.

Given the fact that it costs 5-10 times more to acquire a new customer than to retain an existing one, customer retention has now become even more important than customer acquisition.

For many incumbent operators, retaining high profitable customers is the number one business goal. To reduce customer churn, telecom companies need to predict which customers are at high risk of churn.

Our goal is to build a machine learning model that is able to predict churning customers based on the features provided for their usage.

The steps that we have follwoed below are as follows:

1. Data Understanding, Preparation, Pre-processing
2. Feature Engineering and Variable Transformation
3. Exploratory Data Analysis
4. Model Building
5. Model Recommendations
6. Kaggle Submission

STEP 1: Data Understanding, Preparation, Pre-processing

Data Understanding

Import necessary header files

# Suppress warnings
import warnings
warnings.filterwarnings('ignore')

# Data manipulation libraries
import pandas as pd
import numpy as np
import missingno as msno

# Setting Display Options
pd.set_option("display.max_columns",None)
pd.set_option("display.max_rows",None)
pd.set_option('display.width', None)

# Data visualization libraries
from tabulate import tabulate
import matplotlib.pyplot as plt
import seaborn as sns
import plotly
import plotly.graph_objects as go
import plotly.express as px
import plotly.offline as pyo
from plotly.subplots import make_subplots

# Model building
from sklearn.model_selection import train_test_split
from imblearn.over_sampling import SMOTE
from sklearn.preprocessing import StandardScaler
from sklearn.decomposition import PCA
from sklearn.decomposition import IncrementalPCA
from sklearn.linear_model import LogisticRegression
from sklearn import metrics
from sklearn.metrics import confusion_matrix
from sklearn.metrics import roc_curve, auc
from sklearn.model_selection import KFold
from sklearn.model_selection import StratifiedKFold
from sklearn.model_selection import cross_val_score
from sklearn.model_selection import GridSearchCV
from sklearn.model_selection import RandomizedSearchCV
from sklearn.svm import SVC
from sklearn.tree import DecisionTreeClassifier
from sklearn.ensemble import RandomForestClassifier
import xgboost as xgb
from xgboost import XGBClassifier
from xgboost import plot_importance
import statsmodels.api as sm
from sklearn.linear_model import LogisticRegression
from sklearn.feature_selection import RFE
from statsmodels.stats.outliers_influence import variance_inflation_factor

Load the dataset

telecom = pd.read_csv('train.csv')
telecom.head()

	id	circle_id	loc_og_t2o_mou	std_og_t2o_mou	loc_ic_t2o_mou	last_date_of_month_6	last_date_of_month_7	last_date_of_month_8	arpu_6	arpu_7	arpu_8	onnet_mou_6	onnet_mou_7	onnet_mou_8	offnet_mou_6	offnet_mou_7	offnet_mou_8	roam_ic_mou_6	roam_ic_mou_7	roam_ic_mou_8	roam_og_mou_6	roam_og_mou_7	roam_og_mou_8	loc_og_t2t_mou_6	loc_og_t2t_mou_7	loc_og_t2t_mou_8	loc_og_t2m_mou_6	loc_og_t2m_mou_7	loc_og_t2m_mou_8	loc_og_t2f_mou_6	loc_og_t2f_mou_7	loc_og_t2f_mou_8	loc_og_t2c_mou_6	loc_og_t2c_mou_7	loc_og_t2c_mou_8	loc_og_mou_6	loc_og_mou_7	loc_og_mou_8	std_og_t2t_mou_6	std_og_t2t_mou_7	std_og_t2t_mou_8	std_og_t2m_mou_6	std_og_t2m_mou_7	std_og_t2m_mou_8	std_og_t2f_mou_6	std_og_t2f_mou_7	std_og_t2f_mou_8	std_og_t2c_mou_6	std_og_t2c_mou_7	std_og_t2c_mou_8	std_og_mou_6	std_og_mou_7	std_og_mou_8	isd_og_mou_6	isd_og_mou_7	isd_og_mou_8	spl_og_mou_6	spl_og_mou_7	spl_og_mou_8	og_others_6	og_others_7	og_others_8	total_og_mou_6	total_og_mou_7	total_og_mou_8	loc_ic_t2t_mou_6	loc_ic_t2t_mou_7	loc_ic_t2t_mou_8	loc_ic_t2m_mou_6	loc_ic_t2m_mou_7	loc_ic_t2m_mou_8	loc_ic_t2f_mou_6	loc_ic_t2f_mou_7	loc_ic_t2f_mou_8	loc_ic_mou_6	loc_ic_mou_7	loc_ic_mou_8	std_ic_t2t_mou_6	std_ic_t2t_mou_7	std_ic_t2t_mou_8	std_ic_t2m_mou_6	std_ic_t2m_mou_7	std_ic_t2m_mou_8	std_ic_t2f_mou_6	std_ic_t2f_mou_7	std_ic_t2f_mou_8	std_ic_t2o_mou_6	std_ic_t2o_mou_7	std_ic_t2o_mou_8	std_ic_mou_6	std_ic_mou_7	std_ic_mou_8	total_ic_mou_6	total_ic_mou_7	total_ic_mou_8	spl_ic_mou_6	spl_ic_mou_7	spl_ic_mou_8	isd_ic_mou_6	isd_ic_mou_7	isd_ic_mou_8	ic_others_6	ic_others_7	ic_others_8	total_rech_num_6	total_rech_num_7	total_rech_num_8	total_rech_amt_6	total_rech_amt_7	total_rech_amt_8	max_rech_amt_6	max_rech_amt_7	max_rech_amt_8	date_of_last_rech_6	date_of_last_rech_7	date_of_last_rech_8	last_day_rch_amt_6	last_day_rch_amt_7	last_day_rch_amt_8	date_of_last_rech_data_6	date_of_last_rech_data_7	date_of_last_rech_data_8	total_rech_data_6	total_rech_data_7	total_rech_data_8	max_rech_data_6	max_rech_data_7	max_rech_data_8	count_rech_2g_6	count_rech_2g_7	count_rech_2g_8	count_rech_3g_6	count_rech_3g_7	count_rech_3g_8	av_rech_amt_data_6	av_rech_amt_data_7	av_rech_amt_data_8	vol_2g_mb_6	vol_2g_mb_7	vol_2g_mb_8	vol_3g_mb_6	vol_3g_mb_7	vol_3g_mb_8	arpu_3g_6	arpu_3g_7	arpu_3g_8	arpu_2g_6	arpu_2g_7	arpu_2g_8	night_pck_user_6	night_pck_user_7	night_pck_user_8	monthly_2g_6	monthly_2g_7	monthly_2g_8	sachet_2g_6	sachet_2g_7	sachet_2g_8	monthly_3g_6	monthly_3g_7	monthly_3g_8	sachet_3g_6	sachet_3g_7	sachet_3g_8	fb_user_6	fb_user_7	fb_user_8	aon	aug_vbc_3g	jul_vbc_3g	jun_vbc_3g	churn_probability
0	0	109	0.0	0.0	0.0	6/30/2014	7/31/2014	8/31/2014	31.277	87.009	7.527	48.58	124.38	1.29	32.24	96.68	2.33	0.00	0.0	0.0	0.00	0.0	0.00	2.23	0.00	0.28	5.29	16.04	2.33	0.00	0.00	0.00	0.00	0.00	0.00	7.53	16.04	2.61	46.34	124.38	1.01	18.75	80.61	0.0	0.00	0.0	0.0	0.0	0.0	0.0	65.09	204.99	1.01	0.0	0.0	0.0	8.20	0.63	0.00	0.38	0.0	0.0	81.21	221.68	3.63	2.43	3.68	7.79	0.83	21.08	16.91	0.00	0.00	0.00	3.26	24.76	24.71	0.00	7.61	0.21	7.46	19.96	14.96	0.0	0.0	0.0	0.0	0.0	0.0	7.46	27.58	15.18	11.84	53.04	40.56	0.0	0.0	0.66	0.0	0.0	0.0	1.11	0.69	0.00	3	2	2	77	65	10	65	65	10	6/22/2014	7/10/2014	8/24/2014	65	65	0	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	0.0	0.00	0.00	0.0	0.00	0.0	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	0	0	0	0	0	0	0	0	0	0	0	0	NaN	NaN	NaN	1958	0.0	0.0	0.0	0
1	1	109	0.0	0.0	0.0	6/30/2014	7/31/2014	8/31/2014	0.000	122.787	42.953	0.00	0.00	0.00	0.00	25.99	30.89	0.00	0.0	0.0	0.00	0.0	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	22.01	29.79	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.0	0.00	0.0	0.0	0.0	0.0	0.0	0.00	0.00	0.00	0.0	0.0	0.0	0.00	30.73	31.66	0.00	0.0	0.0	0.00	30.73	31.66	1.68	19.09	10.53	1.41	18.68	11.09	0.35	1.66	3.40	3.44	39.44	25.03	0.00	0.00	0.00	0.00	0.00	0.00	0.0	0.0	0.0	0.0	0.0	0.0	0.00	0.00	0.00	3.44	39.44	25.04	0.0	0.0	0.01	0.0	0.0	0.0	0.00	0.00	0.00	3	4	5	0	145	50	0	145	50	6/12/2014	7/10/2014	8/26/2014	0	0	0	NaN	7/8/2014	NaN	NaN	1.0	NaN	NaN	145.0	NaN	NaN	0.0	NaN	NaN	1.0	NaN	NaN	145.0	NaN	0.0	352.91	0.00	0.0	3.96	0.0	NaN	122.07	NaN	NaN	122.08	NaN	NaN	0.0	NaN	0	0	0	0	0	0	0	1	0	0	0	0	NaN	1.0	NaN	710	0.0	0.0	0.0	0
2	2	109	0.0	0.0	0.0	6/30/2014	7/31/2014	8/31/2014	60.806	103.176	0.000	0.53	15.93	0.00	53.99	82.05	0.00	0.00	0.0	0.0	0.00	0.0	0.00	0.53	12.98	0.00	24.11	0.00	0.00	0.00	0.00	0.00	2.14	0.00	0.00	24.64	12.98	0.00	0.00	2.94	0.00	28.94	82.05	0.0	0.00	0.0	0.0	0.0	0.0	0.0	28.94	84.99	0.00	0.0	0.0	0.0	2.89	1.38	0.00	0.00	0.0	0.0	56.49	99.36	0.00	4.51	6.16	6.49	89.86	25.18	23.51	0.00	0.00	0.00	94.38	31.34	30.01	11.69	0.00	0.00	18.21	2.48	6.38	0.0	0.0	0.0	0.0	0.0	0.0	29.91	2.48	6.38	124.29	33.83	36.64	0.0	0.0	0.00	0.0	0.0	0.0	0.00	0.00	0.25	2	4	2	70	120	0	70	70	0	6/11/2014	7/22/2014	8/24/2014	70	50	0	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	0.0	0.00	0.00	0.0	0.00	0.0	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	0	0	0	0	0	0	0	0	0	0	0	0	NaN	NaN	NaN	882	0.0	0.0	0.0	0
3	3	109	0.0	0.0	0.0	6/30/2014	7/31/2014	8/31/2014	156.362	205.260	111.095	7.26	16.01	0.00	68.76	78.48	50.23	0.00	0.0	0.0	0.00	0.0	1.63	6.99	3.94	0.00	37.91	44.89	23.63	0.00	0.00	0.00	0.00	0.00	8.03	44.91	48.84	23.63	0.26	12.06	0.00	15.33	25.93	4.6	0.56	0.0	0.0	0.0	0.0	0.0	16.16	37.99	4.60	0.0	0.0	0.0	14.95	9.13	25.61	0.00	0.0	0.0	76.03	95.98	53.84	24.98	4.84	23.88	53.99	44.23	57.14	7.23	0.81	0.00	86.21	49.89	81.03	0.00	0.00	0.00	8.89	0.28	2.81	0.0	0.0	0.0	0.0	0.0	0.0	8.89	0.28	2.81	95.11	50.18	83.84	0.0	0.0	0.00	0.0	0.0	0.0	0.00	0.00	0.00	2	4	3	160	240	130	110	110	50	6/15/2014	7/21/2014	8/25/2014	110	110	50	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	0.0	0.00	0.00	0.0	0.00	0.0	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	0	0	0	0	0	0	0	0	0	0	0	0	NaN	NaN	NaN	982	0.0	0.0	0.0	0
4	4	109	0.0	0.0	0.0	6/30/2014	7/31/2014	8/31/2014	240.708	128.191	101.565	21.28	4.83	6.13	56.99	38.11	9.63	53.64	0.0	0.0	15.73	0.0	0.00	10.16	4.83	6.13	36.74	19.88	4.61	11.99	1.23	5.01	0.00	9.85	0.00	58.91	25.94	15.76	0.00	0.00	0.00	4.35	0.00	0.0	0.00	0.0	0.0	0.0	0.0	0.0	4.35	0.00	0.00	0.0	0.0	0.0	0.00	17.00	0.00	0.00	0.0	0.0	63.26	42.94	15.76	5.44	1.39	2.66	10.58	4.33	19.49	5.51	3.63	6.14	21.54	9.36	28.31	0.00	0.00	0.00	0.00	0.00	0.00	0.0	0.0	0.0	0.0	0.0	0.0	0.00	0.00	0.00	21.54	9.36	28.31	0.0	0.0	0.00	0.0	0.0	0.0	0.00	0.00	0.00	13	10	8	290	136	122	50	41	30	6/25/2014	7/26/2014	8/30/2014	25	10	30	6/25/2014	7/23/2014	8/20/2014	7.0	7.0	6.0	25.0	41.0	25.0	7.0	6.0	6.0	0.0	1.0	0.0	175.0	191.0	142.0	390.8	308.89	213.47	0.0	0.00	0.0	0.0	35.00	0.0	0.0	35.12	0.0	0.0	0.0	0.0	0	0	0	7	6	6	0	0	0	0	1	0	1.0	1.0	1.0	647	0.0	0.0	0.0	0

Data Overview

# How many rows and columns are present?
telecom.shape

(69999, 172)

The dataset has 172 columns and 69999 rows of data. Since we have 172 columns, let's remove columns that we do not need and then go ahead and further understand the data.

# What are the datatypes involved?
telecom.info(verbose=1)

<class 'pandas.core.frame.DataFrame'>
RangeIndex: 69999 entries, 0 to 69998
Data columns (total 172 columns):
 #    Column                    Dtype  
---   ------                    -----  
 0    id                        int64  
 1    circle_id                 int64  
 2    loc_og_t2o_mou            float64
 3    std_og_t2o_mou            float64
 4    loc_ic_t2o_mou            float64
 5    last_date_of_month_6      object 
 6    last_date_of_month_7      object 
 7    last_date_of_month_8      object 
 8    arpu_6                    float64
 9    arpu_7                    float64
 10   arpu_8                    float64
 11   onnet_mou_6               float64
 12   onnet_mou_7               float64
 13   onnet_mou_8               float64
 14   offnet_mou_6              float64
 15   offnet_mou_7              float64
 16   offnet_mou_8              float64
 17   roam_ic_mou_6             float64
 18   roam_ic_mou_7             float64
 19   roam_ic_mou_8             float64
 20   roam_og_mou_6             float64
 21   roam_og_mou_7             float64
 22   roam_og_mou_8             float64
 23   loc_og_t2t_mou_6          float64
 24   loc_og_t2t_mou_7          float64
 25   loc_og_t2t_mou_8          float64
 26   loc_og_t2m_mou_6          float64
 27   loc_og_t2m_mou_7          float64
 28   loc_og_t2m_mou_8          float64
 29   loc_og_t2f_mou_6          float64
 30   loc_og_t2f_mou_7          float64
 31   loc_og_t2f_mou_8          float64
 32   loc_og_t2c_mou_6          float64
 33   loc_og_t2c_mou_7          float64
 34   loc_og_t2c_mou_8          float64
 35   loc_og_mou_6              float64
 36   loc_og_mou_7              float64
 37   loc_og_mou_8              float64
 38   std_og_t2t_mou_6          float64
 39   std_og_t2t_mou_7          float64
 40   std_og_t2t_mou_8          float64
 41   std_og_t2m_mou_6          float64
 42   std_og_t2m_mou_7          float64
 43   std_og_t2m_mou_8          float64
 44   std_og_t2f_mou_6          float64
 45   std_og_t2f_mou_7          float64
 46   std_og_t2f_mou_8          float64
 47   std_og_t2c_mou_6          float64
 48   std_og_t2c_mou_7          float64
 49   std_og_t2c_mou_8          float64
 50   std_og_mou_6              float64
 51   std_og_mou_7              float64
 52   std_og_mou_8              float64
 53   isd_og_mou_6              float64
 54   isd_og_mou_7              float64
 55   isd_og_mou_8              float64
 56   spl_og_mou_6              float64
 57   spl_og_mou_7              float64
 58   spl_og_mou_8              float64
 59   og_others_6               float64
 60   og_others_7               float64
 61   og_others_8               float64
 62   total_og_mou_6            float64
 63   total_og_mou_7            float64
 64   total_og_mou_8            float64
 65   loc_ic_t2t_mou_6          float64
 66   loc_ic_t2t_mou_7          float64
 67   loc_ic_t2t_mou_8          float64
 68   loc_ic_t2m_mou_6          float64
 69   loc_ic_t2m_mou_7          float64
 70   loc_ic_t2m_mou_8          float64
 71   loc_ic_t2f_mou_6          float64
 72   loc_ic_t2f_mou_7          float64
 73   loc_ic_t2f_mou_8          float64
 74   loc_ic_mou_6              float64
 75   loc_ic_mou_7              float64
 76   loc_ic_mou_8              float64
 77   std_ic_t2t_mou_6          float64
 78   std_ic_t2t_mou_7          float64
 79   std_ic_t2t_mou_8          float64
 80   std_ic_t2m_mou_6          float64
 81   std_ic_t2m_mou_7          float64
 82   std_ic_t2m_mou_8          float64
 83   std_ic_t2f_mou_6          float64
 84   std_ic_t2f_mou_7          float64
 85   std_ic_t2f_mou_8          float64
 86   std_ic_t2o_mou_6          float64
 87   std_ic_t2o_mou_7          float64
 88   std_ic_t2o_mou_8          float64
 89   std_ic_mou_6              float64
 90   std_ic_mou_7              float64
 91   std_ic_mou_8              float64
 92   total_ic_mou_6            float64
 93   total_ic_mou_7            float64
 94   total_ic_mou_8            float64
 95   spl_ic_mou_6              float64
 96   spl_ic_mou_7              float64
 97   spl_ic_mou_8              float64
 98   isd_ic_mou_6              float64
 99   isd_ic_mou_7              float64
 100  isd_ic_mou_8              float64
 101  ic_others_6               float64
 102  ic_others_7               float64
 103  ic_others_8               float64
 104  total_rech_num_6          int64  
 105  total_rech_num_7          int64  
 106  total_rech_num_8          int64  
 107  total_rech_amt_6          int64  
 108  total_rech_amt_7          int64  
 109  total_rech_amt_8          int64  
 110  max_rech_amt_6            int64  
 111  max_rech_amt_7            int64  
 112  max_rech_amt_8            int64  
 113  date_of_last_rech_6       object 
 114  date_of_last_rech_7       object 
 115  date_of_last_rech_8       object 
 116  last_day_rch_amt_6        int64  
 117  last_day_rch_amt_7        int64  
 118  last_day_rch_amt_8        int64  
 119  date_of_last_rech_data_6  object 
 120  date_of_last_rech_data_7  object 
 121  date_of_last_rech_data_8  object 
 122  total_rech_data_6         float64
 123  total_rech_data_7         float64
 124  total_rech_data_8         float64
 125  max_rech_data_6           float64
 126  max_rech_data_7           float64
 127  max_rech_data_8           float64
 128  count_rech_2g_6           float64
 129  count_rech_2g_7           float64
 130  count_rech_2g_8           float64
 131  count_rech_3g_6           float64
 132  count_rech_3g_7           float64
 133  count_rech_3g_8           float64
 134  av_rech_amt_data_6        float64
 135  av_rech_amt_data_7        float64
 136  av_rech_amt_data_8        float64
 137  vol_2g_mb_6               float64
 138  vol_2g_mb_7               float64
 139  vol_2g_mb_8               float64
 140  vol_3g_mb_6               float64
 141  vol_3g_mb_7               float64
 142  vol_3g_mb_8               float64
 143  arpu_3g_6                 float64
 144  arpu_3g_7                 float64
 145  arpu_3g_8                 float64
 146  arpu_2g_6                 float64
 147  arpu_2g_7                 float64
 148  arpu_2g_8                 float64
 149  night_pck_user_6          float64
 150  night_pck_user_7          float64
 151  night_pck_user_8          float64
 152  monthly_2g_6              int64  
 153  monthly_2g_7              int64  
 154  monthly_2g_8              int64  
 155  sachet_2g_6               int64  
 156  sachet_2g_7               int64  
 157  sachet_2g_8               int64  
 158  monthly_3g_6              int64  
 159  monthly_3g_7              int64  
 160  monthly_3g_8              int64  
 161  sachet_3g_6               int64  
 162  sachet_3g_7               int64  
 163  sachet_3g_8               int64  
 164  fb_user_6                 float64
 165  fb_user_7                 float64
 166  fb_user_8                 float64
 167  aon                       int64  
 168  aug_vbc_3g                float64
 169  jul_vbc_3g                float64
 170  jun_vbc_3g                float64
 171  churn_probability         int64  
dtypes: float64(135), int64(28), object(9)
memory usage: 91.9+ MB

# Check the dataspread
telecom.describe()

	id	circle_id	loc_og_t2o_mou	std_og_t2o_mou	loc_ic_t2o_mou	arpu_6	arpu_7	arpu_8	onnet_mou_6	onnet_mou_7	onnet_mou_8	offnet_mou_6	offnet_mou_7	offnet_mou_8	roam_ic_mou_6	roam_ic_mou_7	roam_ic_mou_8	roam_og_mou_6	roam_og_mou_7	roam_og_mou_8	loc_og_t2t_mou_6	loc_og_t2t_mou_7	loc_og_t2t_mou_8	loc_og_t2m_mou_6	loc_og_t2m_mou_7	loc_og_t2m_mou_8	loc_og_t2f_mou_6	loc_og_t2f_mou_7	loc_og_t2f_mou_8	loc_og_t2c_mou_6	loc_og_t2c_mou_7	loc_og_t2c_mou_8	loc_og_mou_6	loc_og_mou_7	loc_og_mou_8	std_og_t2t_mou_6	std_og_t2t_mou_7	std_og_t2t_mou_8	std_og_t2m_mou_6	std_og_t2m_mou_7	std_og_t2m_mou_8	std_og_t2f_mou_6	std_og_t2f_mou_7	std_og_t2f_mou_8	std_og_t2c_mou_6	std_og_t2c_mou_7	std_og_t2c_mou_8	std_og_mou_6	std_og_mou_7	std_og_mou_8	isd_og_mou_6	isd_og_mou_7	isd_og_mou_8	spl_og_mou_6	spl_og_mou_7	spl_og_mou_8	og_others_6	og_others_7	og_others_8	total_og_mou_6	total_og_mou_7	total_og_mou_8	loc_ic_t2t_mou_6	loc_ic_t2t_mou_7	loc_ic_t2t_mou_8	loc_ic_t2m_mou_6	loc_ic_t2m_mou_7	loc_ic_t2m_mou_8	loc_ic_t2f_mou_6	loc_ic_t2f_mou_7	loc_ic_t2f_mou_8	loc_ic_mou_6	loc_ic_mou_7	loc_ic_mou_8	std_ic_t2t_mou_6	std_ic_t2t_mou_7	std_ic_t2t_mou_8	std_ic_t2m_mou_6	std_ic_t2m_mou_7	std_ic_t2m_mou_8	std_ic_t2f_mou_6	std_ic_t2f_mou_7	std_ic_t2f_mou_8	std_ic_t2o_mou_6	std_ic_t2o_mou_7	std_ic_t2o_mou_8	std_ic_mou_6	std_ic_mou_7	std_ic_mou_8	total_ic_mou_6	total_ic_mou_7	total_ic_mou_8	spl_ic_mou_6	spl_ic_mou_7	spl_ic_mou_8	isd_ic_mou_6	isd_ic_mou_7	isd_ic_mou_8	ic_others_6	ic_others_7	ic_others_8	total_rech_num_6	total_rech_num_7	total_rech_num_8	total_rech_amt_6	total_rech_amt_7	total_rech_amt_8	max_rech_amt_6	max_rech_amt_7	max_rech_amt_8	last_day_rch_amt_6	last_day_rch_amt_7	last_day_rch_amt_8	total_rech_data_6	total_rech_data_7	total_rech_data_8	max_rech_data_6	max_rech_data_7	max_rech_data_8	count_rech_2g_6	count_rech_2g_7	count_rech_2g_8	count_rech_3g_6	count_rech_3g_7	count_rech_3g_8	av_rech_amt_data_6	av_rech_amt_data_7	av_rech_amt_data_8	vol_2g_mb_6	vol_2g_mb_7	vol_2g_mb_8	vol_3g_mb_6	vol_3g_mb_7	vol_3g_mb_8	arpu_3g_6	arpu_3g_7	arpu_3g_8	arpu_2g_6	arpu_2g_7	arpu_2g_8	night_pck_user_6	night_pck_user_7	night_pck_user_8	monthly_2g_6	monthly_2g_7	monthly_2g_8	sachet_2g_6	sachet_2g_7	sachet_2g_8	monthly_3g_6	monthly_3g_7	monthly_3g_8	sachet_3g_6	sachet_3g_7	sachet_3g_8	fb_user_6	fb_user_7	fb_user_8	aon	aug_vbc_3g	jul_vbc_3g	jun_vbc_3g	churn_probability
count	69999.000000	69999.0	69297.0	69297.0	69297.0	69999.000000	69999.000000	69999.000000	67231.000000	67312.000000	66296.000000	67231.000000	67312.000000	66296.000000	67231.000000	67312.000000	66296.000000	67231.000000	67312.000000	66296.000000	67231.000000	67312.000000	66296.000000	67231.000000	67312.000000	66296.000000	67231.000000	67312.000000	66296.000000	67231.000000	67312.000000	66296.000000	67231.000000	67312.000000	66296.000000	67231.000000	67312.000000	66296.000000	67231.000000	67312.000000	66296.000000	67231.000000	67312.000000	66296.000000	67231.0	67312.0	66296.0	67231.000000	67312.000000	66296.000000	67231.000000	67312.000000	66296.000000	67231.000000	67312.000000	66296.000000	67231.000000	67312.000000	66296.000000	69999.000000	69999.000000	69999.000000	67231.000000	67312.000000	66296.000000	67231.000000	67312.000000	66296.000000	67231.000000	67312.000000	66296.000000	67231.000000	67312.000000	66296.000000	67231.000000	67312.000000	66296.000000	67231.000000	67312.000000	66296.000000	67231.000000	67312.000000	66296.000000	67231.0	67312.0	66296.0	67231.000000	67312.000000	66296.000000	69999.000000	69999.000000	69999.000000	67231.000000	67312.000000	66296.000000	67231.000000	67312.000000	66296.000000	67231.000000	67312.000000	66296.000000	69999.000000	69999.000000	69999.000000	69999.000000	69999.000000	69999.000000	69999.000000	69999.000000	69999.000000	69999.000000	69999.000000	69999.000000	17568.000000	17865.000000	18417.000000	17568.000000	17865.000000	18417.000000	17568.000000	17865.000000	18417.000000	17568.000000	17865.000000	18417.000000	17568.000000	17865.000000	18417.000000	69999.000000	69999.000000	69999.000000	69999.000000	69999.000000	69999.000000	17568.000000	17865.000000	18417.000000	17568.000000	17865.000000	18417.000000	17568.000000	17865.000000	18417.000000	69999.000000	69999.000000	69999.000000	69999.000000	69999.000000	69999.000000	69999.000000	69999.000000	69999.000000	69999.000000	69999.000000	69999.000000	17568.000000	17865.000000	18417.000000	69999.000000	69999.000000	69999.000000	69999.00000	69999.000000
mean	34999.000000	109.0	0.0	0.0	0.0	283.134365	278.185912	278.858826	133.153275	133.894438	132.978257	198.874771	197.153383	196.543577	9.765435	7.014568	7.004892	14.186457	9.842191	9.771783	46.904854	46.166503	45.686109	93.238231	90.799240	91.121447	3.743179	3.777031	3.661652	1.126025	1.361052	1.420840	143.893585	140.750120	140.476486	80.619382	83.775851	83.471486	88.152110	91.538615	90.586999	1.126377	1.084062	1.057739	0.0	0.0	0.0	169.900601	176.401217	175.118852	0.845763	0.811100	0.841648	3.958619	4.976783	5.045027	0.462581	0.024425	0.033059	306.451436	310.572674	304.513065	48.043255	47.882736	47.256388	107.152439	106.489856	108.154731	12.050672	12.563665	11.716763	167.255126	166.945103	167.136761	9.476958	9.873468	9.910217	20.734858	21.685359	21.089042	2.146273	2.199395	2.075179	0.0	0.0	0.0	32.360632	33.760809	33.077030	199.710640	201.878029	198.486034	0.061932	0.033371	0.040392	7.394167	8.171162	8.348424	0.854063	1.019680	0.963214	7.566522	7.706667	7.224932	328.139788	322.376363	323.846355	104.569265	104.137573	107.540351	63.426949	59.294218	62.489478	2.467612	2.679989	2.652441	126.500000	126.402071	125.374925	1.865323	2.056311	2.016018	0.602288	0.623678	0.636423	192.831096	201.455940	196.815792	51.773924	51.240204	50.127506	122.171882	128.934444	135.486541	90.069931	89.115767	90.618564	86.863900	85.846074	86.348404	0.025273	0.024069	0.021013	0.079287	0.083401	0.080930	0.388863	0.441406	0.449492	0.075815	0.077730	0.081958	0.075344	0.081444	0.085487	0.916325	0.909544	0.890319	1220.639709	68.108597	65.935830	60.07674	0.101887
std	20207.115084	0.0	0.0	0.0	0.0	334.213918	344.366927	351.924315	299.963093	311.277193	311.896596	316.818355	322.482226	324.089234	57.374429	55.960985	53.408135	73.469261	58.511894	64.618388	150.971758	154.739002	153.716880	162.046699	153.852597	152.997805	13.319542	13.568110	13.009193	5.741811	7.914113	6.542202	252.034597	246.313148	245.342359	255.098355	266.693254	267.021929	255.771554	267.532089	270.032002	8.136645	8.325206	7.696853	0.0	0.0	0.0	392.046600	409.299501	410.697098	29.747486	29.220073	29.563367	15.854529	22.229842	17.708507	4.768437	1.716430	2.232547	465.502866	479.131770	477.936832	140.499757	147.761124	141.249368	168.455999	165.452459	166.223461	39.416076	43.495179	38.606895	252.576231	254.688718	249.288410	51.664472	56.137824	54.248186	80.294236	87.314510	81.534344	16.522232	16.171533	15.865403	0.0	0.0	0.0	104.381082	114.142230	108.469864	290.114823	296.771338	288.336731	0.164823	0.137322	0.148417	60.951165	63.604165	63.097570	12.149144	13.225373	11.697686	7.041452	7.050614	7.195597	404.211068	411.070120	426.181405	121.407701	120.782543	124.396750	97.954876	95.429492	101.996729	2.794610	3.073472	3.101265	109.352573	109.459266	109.648799	2.566377	2.799916	2.728246	1.279297	1.401230	1.457058	190.623115	198.346141	192.280532	212.513909	211.114667	213.101403	554.869965	554.096072	568.310234	193.600413	195.826990	189.907986	171.321203	178.067280	170.297094	0.156958	0.153269	0.143432	0.294719	0.304802	0.299254	1.494206	1.651012	1.632450	0.358905	0.383189	0.381821	0.573003	0.634547	0.680035	0.276907	0.286842	0.312501	952.426321	269.328659	267.899034	257.22681	0.302502
min	0.000000	109.0	0.0	0.0	0.0	-2258.709000	-1289.715000	-945.808000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.0	0.0	0.0	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.0	0.0	0.0	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	1.000000	1.000000	1.000000	1.000000	1.000000	1.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	1.000000	1.000000	1.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	-20.380000	-26.040000	-24.490000	-35.830000	-13.090000	-55.830000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	180.000000	0.000000	0.000000	0.00000	0.000000
25%	17499.500000	109.0	0.0	0.0	0.0	93.581000	86.714000	84.095000	7.410000	6.675000	6.410000	34.860000	32.240000	31.575000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	1.660000	1.650000	1.610000	9.920000	10.090000	9.830000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	17.235000	17.590000	17.237500	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.0	0.0	0.0	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	44.780000	42.910000	38.710000	3.030000	3.260000	3.280000	17.390000	18.610000	18.940000	0.000000	0.000000	0.000000	30.630000	32.710000	32.810000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.0	0.0	0.0	0.000000	0.000000	0.030000	38.640000	41.340000	38.290000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	3.000000	3.000000	3.000000	110.000000	100.000000	90.000000	30.000000	30.000000	30.000000	0.000000	0.000000	0.000000	1.000000	1.000000	1.000000	25.000000	25.000000	25.000000	1.000000	1.000000	1.000000	0.000000	0.000000	0.000000	82.000000	92.000000	84.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	1.000000	1.000000	1.000000	468.000000	0.000000	0.000000	0.00000	0.000000
50%	34999.000000	109.0	0.0	0.0	0.0	197.484000	191.588000	192.234000	34.110000	32.280000	32.100000	96.480000	91.885000	91.800000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	11.910000	11.580000	11.740000	41.030000	40.170000	40.350000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	65.190000	63.430000	63.520000	0.000000	0.000000	0.000000	3.980000	3.710000	3.300000	0.000000	0.000000	0.000000	0.0	0.0	0.0	11.730000	11.260000	10.505000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	145.280000	141.230000	138.360000	15.740000	15.830000	16.040000	56.460000	56.930000	58.210000	0.880000	0.910000	0.930000	92.430000	92.510000	93.890000	0.000000	0.000000	0.000000	2.040000	2.060000	2.030000	0.000000	0.000000	0.000000	0.0	0.0	0.0	5.910000	5.980000	5.830000	114.780000	116.330000	114.610000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	6.000000	6.000000	5.000000	229.000000	220.000000	225.000000	110.000000	110.000000	98.000000	30.000000	30.000000	30.000000	1.000000	2.000000	1.000000	145.000000	145.000000	145.000000	1.000000	1.000000	1.000000	0.000000	0.000000	0.000000	154.000000	154.000000	154.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.520000	0.420000	0.840000	11.300000	8.800000	9.090000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	1.000000	1.000000	1.000000	868.000000	0.000000	0.000000	0.00000	0.000000
75%	52498.500000	109.0	0.0	0.0	0.0	370.791000	365.369500	369.909000	119.390000	115.837500	115.060000	232.990000	227.630000	229.345000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	40.740000	39.760000	39.895000	110.430000	107.540000	109.245000	2.060000	2.080000	2.030000	0.000000	0.000000	0.000000	167.880000	163.932500	165.615000	31.020000	31.300000	30.760000	53.745000	54.640000	52.660000	0.000000	0.000000	0.000000	0.0	0.0	0.0	146.335000	151.645000	149.015000	0.000000	0.000000	0.000000	2.400000	3.660000	4.002500	0.000000	0.000000	0.000000	374.305000	380.045000	370.895000	46.980000	45.690000	46.280000	132.020000	131.010000	134.380000	8.140000	8.230000	8.090000	208.325000	205.530000	208.060000	4.060000	4.180000	4.052500	14.960000	15.830000	15.310000	0.000000	0.000000	0.000000	0.0	0.0	0.0	26.780000	28.160000	27.615000	251.070000	249.470000	249.710000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	9.000000	10.000000	9.000000	438.000000	430.000000	436.000000	120.000000	128.000000	144.000000	110.000000	110.000000	130.000000	3.000000	3.000000	3.000000	177.000000	177.000000	179.000000	2.000000	2.000000	2.000000	1.000000	1.000000	1.000000	252.000000	252.000000	252.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	122.070000	120.860000	122.070000	122.070000	122.070000	122.070000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	1.000000	1.000000	1.000000	1813.000000	0.000000	0.000000	0.00000	0.000000
max	69998.000000	109.0	0.0	0.0	0.0	27731.088000	35145.834000	33543.624000	7376.710000	8157.780000	10752.560000	8362.360000	7043.980000	14007.340000	2850.980000	4155.830000	4169.810000	3775.110000	2812.040000	5337.040000	6431.330000	7400.660000	10752.560000	4696.830000	4557.140000	4961.330000	617.580000	815.330000	588.290000	342.860000	916.240000	351.830000	10643.380000	7674.780000	11039.910000	7366.580000	8133.660000	8014.430000	8314.760000	6622.540000	13950.040000	628.560000	465.790000	354.160000	0.0	0.0	0.0	8432.990000	8155.530000	13980.060000	5900.660000	5490.280000	5681.540000	1023.210000	2372.510000	1075.080000	800.890000	270.240000	394.930000	10674.030000	8285.640000	14043.060000	5315.590000	9324.660000	10696.230000	4450.740000	4455.830000	6274.190000	1872.340000	1983.010000	1676.580000	7454.630000	9669.910000	10830.160000	3336.380000	4708.710000	3930.240000	5647.160000	6141.880000	5512.760000	1351.110000	1136.080000	1394.890000	0.0	0.0	0.0	5712.110000	6745.760000	5658.740000	7716.140000	9699.010000	10830.380000	19.760000	13.460000	16.860000	6789.410000	5289.540000	4127.010000	1362.940000	1495.940000	1209.860000	170.000000	138.000000	138.000000	35190.000000	40335.000000	45320.000000	4010.000000	3299.000000	4449.000000	4010.000000	3100.000000	4449.000000	61.000000	54.000000	60.000000	1555.000000	1555.000000	1555.000000	42.000000	48.000000	44.000000	29.000000	34.000000	45.000000	5920.000000	4365.000000	4076.000000	10285.900000	7873.550000	11117.610000	45735.400000	28144.120000	30036.060000	5054.370000	4980.900000	3716.900000	5054.350000	4809.360000	3483.170000	1.000000	1.000000	1.000000	4.000000	5.000000	5.000000	42.000000	48.000000	44.000000	9.000000	16.000000	16.000000	29.000000	33.000000	41.000000	1.000000	1.000000	1.000000	4337.000000	12916.220000	9165.600000	11166.21000	1.000000

A lot of outliers, negative values and different scales of data can be observed.

Data Preparation and Pre-processing

Rename columns for uniform processing

# Recoding month abbreviations to numbers so that all columns representing months are uniformly named
months = ['aug_vbc_3g','jul_vbc_3g','jun_vbc_3g']
telecom = telecom.rename(columns = {'aug_vbc_3g':'3g_vbc_8','jul_vbc_3g':'3g_vbc_7','jun_vbc_3g':'3g_vbc_6'})

Filter high value customers

We define high value customers as those customers whose avergae total recharge amount is in the 70th percentile.

#Impute null with 0 - indicates recharge was not done

telecom['total_rech_data_6'] = telecom['total_rech_data_6'].replace(np.NaN,0.0)
telecom['total_rech_data_7'] = telecom['total_rech_data_7'].replace(np.NaN,0.0)
telecom['total_rech_data_8'] = telecom['total_rech_data_8'].replace(np.NaN,0.0)
telecom['av_rech_amt_data_6'] = telecom['av_rech_amt_data_6'].replace(np.NaN,0.0)
telecom['av_rech_amt_data_7'] = telecom['av_rech_amt_data_7'].replace(np.NaN,0.0)
telecom['av_rech_amt_data_8'] = telecom['av_rech_amt_data_8'].replace(np.NaN,0.0)

# Create new column: total recharge amount for data: total_rech_amt_data 
telecom['total_rech_amt_data_6'] = telecom.av_rech_amt_data_6 * telecom.total_rech_data_6
telecom['total_rech_amt_data_7'] = telecom.av_rech_amt_data_7 * telecom.total_rech_data_7
telecom['total_rech_amt_data_8'] = telecom.av_rech_amt_data_8 * telecom.total_rech_data_8

# Create total average recharge amount for months 6 and 7

telecom['total_avg_rech_amt_6_7'] = (telecom.total_rech_amt_6 + telecom.total_rech_amt_data_6 \
                                               + telecom.total_rech_amt_7+ telecom.total_rech_amt_data_7)/2

# High value customers have recharge amount in the 70th percentile
high_value_filter = telecom.total_avg_rech_amt_6_7.quantile(0.7)

print('70 percentile of 6th and 7th months avg recharge amount: '+str(high_value_filter))

telecom_high_val_cust = telecom[telecom.total_avg_rech_amt_6_7 > high_value_filter]
print('Dataframe shape after filtering High Value Customers: ' + str(telecom_high_val_cust.shape))

70 percentile of 6th and 7th months avg recharge amount: 477.5
Dataframe shape after filtering High Value Customers: (20998, 176)

# New column total_rech_amt_data_* is created.
# Drop av_rech_amt_data_* and total_rech_data_*
columns_to_drop = ['av_rech_amt_data_6','av_rech_amt_data_7','av_rech_amt_data_8',
                  'total_rech_data_6','total_rech_data_7','total_rech_data_8']
telecom_high_val_cust = telecom_high_val_cust.drop(columns_to_drop, axis=1)
print("Dataframe shape after dropping transformed columns:",str(telecom_high_val_cust.shape))

Dataframe shape after dropping transformed columns: (20998, 170)

Drop columns with single value

# Columns with single values do not add any value to data analysis or modeling; therefore, drop them.

biased_columns = []
for column in telecom_high_val_cust.columns:
    if telecom_high_val_cust[column].nunique() ==1:
        biased_columns.append(column)

print("Number of columns with biased data:",len(biased_columns))
print("We will drop them.")

# Dropping columns that have only one unique value
telecom_high_val_cust = telecom_high_val_cust.drop(biased_columns,axis=1)
print("Dataframe shape after dropping biased columns:",str(telecom_high_val_cust.shape))

Number of columns with biased data: 13
We will drop them.
Dataframe shape after dropping biased columns: (20998, 157)

Drop columns that are not useful for analysis

#date columns do not add any value to our analysis
date_columns = [col for col in telecom_high_val_cust.columns if 'date' in col]
telecom_high_val_cust = telecom_high_val_cust.drop(date_columns, axis=1)

# no analysis will be done on id
telecom_high_val_cust = telecom_high_val_cust.drop('id', axis=1)

# drop column created to identify high value customers
telecom_high_val_cust = telecom_high_val_cust.drop('total_avg_rech_amt_6_7', axis=1)

print("Dataframe shape after dropping columns not useful for analysis:",str(telecom_high_val_cust.shape))

Dataframe shape after dropping columns not useful for analysis: (20998, 149)

Analyse missing data in columns

# Using missingno library, identify columns with missing values
msno.bar(telecom_high_val_cust)

<Axes: >

Dataset metadata

def metadata_matrix(data) :
    return pd.DataFrame({
                'Datatype' : telecom_high_val_cust.dtypes.astype(str),
                'Null_Percentage': round(telecom_high_val_cust.isnull().sum()/len(telecom_high_val_cust) * 100 , 2),
                'Unique_Values_Count': telecom_high_val_cust.nunique().astype(int)
                 }).sort_values(by='Null_Percentage', ascending=False)

metadata_matrix(telecom_high_val_cust)

	Datatype	Null_Percentage	Unique_Values_Count
arpu_3g_8	float64	46.78	4412
fb_user_8	float64	46.78	2
arpu_2g_8	float64	46.78	3604
max_rech_data_8	float64	46.78	46
night_pck_user_8	float64	46.78	2
count_rech_3g_8	float64	46.78	29
count_rech_2g_8	float64	46.78	33
arpu_2g_6	float64	44.35	4363
count_rech_3g_6	float64	44.35	23
max_rech_data_6	float64	44.35	46
count_rech_2g_6	float64	44.35	29
fb_user_6	float64	44.35	2
night_pck_user_6	float64	44.35	2
arpu_3g_6	float64	44.35	4627
count_rech_3g_7	float64	43.26	27
night_pck_user_7	float64	43.26	2
max_rech_data_7	float64	43.26	44
fb_user_7	float64	43.26	2
arpu_3g_7	float64	43.26	4572
count_rech_2g_7	float64	43.26	33
arpu_2g_7	float64	43.26	4102
loc_ic_t2f_mou_8	float64	3.91	3933
isd_og_mou_8	float64	3.91	690
spl_og_mou_8	float64	3.91	2779
og_others_8	float64	3.91	103
loc_ic_t2t_mou_8	float64	3.91	7801
loc_ic_t2m_mou_8	float64	3.91	12147
std_ic_mou_8	float64	3.91	6382
loc_ic_mou_8	float64	3.91	14101
std_ic_t2t_mou_8	float64	3.91	3517
std_ic_t2m_mou_8	float64	3.91	5071
spl_ic_mou_8	float64	3.91	79
isd_ic_mou_8	float64	3.91	2621
ic_others_8	float64	3.91	1020
std_og_t2f_mou_8	float64	3.91	1221
std_og_mou_8	float64	3.91	10986
std_ic_t2f_mou_8	float64	3.91	1579
loc_og_mou_8	float64	3.91	13524
loc_og_t2f_mou_8	float64	3.91	2526
std_og_t2m_mou_8	float64	3.91	9046
onnet_mou_8	float64	3.91	12309
loc_og_t2t_mou_8	float64	3.91	8179
roam_ic_mou_8	float64	3.91	2598
loc_og_t2c_mou_8	float64	3.91	1450
loc_og_t2m_mou_8	float64	3.91	11978
roam_og_mou_8	float64	3.91	3001
std_og_t2t_mou_8	float64	3.91	7749
offnet_mou_8	float64	3.91	15333
std_ic_t2m_mou_6	float64	1.89	5282
loc_ic_mou_6	float64	1.89	14454
loc_og_mou_6	float64	1.89	14173
std_ic_t2t_mou_6	float64	1.89	3609
std_ic_t2f_mou_6	float64	1.89	1608
loc_og_t2m_mou_6	float64	1.89	12396
std_ic_mou_6	float64	1.89	6631
ic_others_6	float64	1.89	989
loc_og_t2c_mou_6	float64	1.89	1432
loc_og_t2f_mou_6	float64	1.89	2609
std_og_t2f_mou_6	float64	1.89	1322
spl_ic_mou_6	float64	1.89	74
loc_og_t2t_mou_6	float64	1.89	8510
loc_ic_t2f_mou_6	float64	1.89	3994
isd_ic_mou_6	float64	1.89	2555
offnet_mou_6	float64	1.89	16091
std_og_mou_6	float64	1.89	11930
onnet_mou_6	float64	1.89	13084
loc_ic_t2m_mou_6	float64	1.89	12503
roam_og_mou_6	float64	1.89	3559
loc_ic_t2t_mou_6	float64	1.89	7960
isd_og_mou_6	float64	1.89	792
roam_ic_mou_6	float64	1.89	3134
std_og_t2t_mou_6	float64	1.89	8384
og_others_6	float64	1.89	755
spl_og_mou_6	float64	1.89	2632
std_og_t2m_mou_6	float64	1.89	9908
onnet_mou_7	float64	1.87	13118
offnet_mou_7	float64	1.87	16126
ic_others_7	float64	1.87	1107
loc_og_t2m_mou_7	float64	1.87	12364
roam_ic_mou_7	float64	1.87	2616
roam_og_mou_7	float64	1.87	2966
loc_og_t2t_mou_7	float64	1.87	8476
isd_ic_mou_7	float64	1.87	2686
std_og_t2f_mou_7	float64	1.87	1265
loc_og_t2f_mou_7	float64	1.87	2622
spl_ic_mou_7	float64	1.87	83
std_og_mou_7	float64	1.87	11861
isd_og_mou_7	float64	1.87	791
std_og_t2m_mou_7	float64	1.87	9878
spl_og_mou_7	float64	1.87	2843
og_others_7	float64	1.87	87
loc_ic_t2t_mou_7	float64	1.87	8021
std_og_t2t_mou_7	float64	1.87	8466
loc_ic_t2m_mou_7	float64	1.87	12503
loc_ic_t2f_mou_7	float64	1.87	4039
loc_og_mou_7	float64	1.87	14194
loc_ic_mou_7	float64	1.87	14495
std_ic_t2t_mou_7	float64	1.87	3703
std_ic_mou_7	float64	1.87	6703
std_ic_t2f_mou_7	float64	1.87	1666
loc_og_t2c_mou_7	float64	1.87	1479
std_ic_t2m_mou_7	float64	1.87	5391
sachet_2g_8	int64	0.00	32
churn_probability	int64	0.00	2
total_rech_amt_data_7	float64	0.00	1448
3g_vbc_8	float64	0.00	6959
3g_vbc_7	float64	0.00	7100
total_rech_amt_data_6	float64	0.00	1357
monthly_2g_6	int64	0.00	5
monthly_2g_7	int64	0.00	6
3g_vbc_6	float64	0.00	6599
sachet_2g_7	int64	0.00	33
sachet_3g_8	int64	0.00	28
sachet_3g_7	int64	0.00	26
sachet_3g_6	int64	0.00	23
monthly_3g_8	int64	0.00	12
monthly_3g_7	int64	0.00	13
monthly_3g_6	int64	0.00	10
monthly_2g_8	int64	0.00	6
sachet_2g_6	int64	0.00	29
aon	int64	0.00	3139
arpu_6	float64	0.00	20453
vol_3g_mb_8	float64	0.00	7157
total_rech_amt_6	int64	0.00	2093
arpu_8	float64	0.00	19775
total_og_mou_6	float64	0.00	17265
total_og_mou_7	float64	0.00	17392
total_og_mou_8	float64	0.00	16445
arpu_7	float64	0.00	20461
total_ic_mou_6	float64	0.00	15440
total_ic_mou_7	float64	0.00	15480
total_ic_mou_8	float64	0.00	15089
total_rech_num_6	int64	0.00	94
total_rech_num_7	int64	0.00	96
total_rech_num_8	int64	0.00	90
total_rech_amt_7	int64	0.00	2129
vol_3g_mb_7	float64	0.00	7631
total_rech_amt_8	int64	0.00	2104
max_rech_amt_6	int64	0.00	159
max_rech_amt_7	int64	0.00	153
max_rech_amt_8	int64	0.00	170
last_day_rch_amt_6	int64	0.00	150
last_day_rch_amt_7	int64	0.00	145
last_day_rch_amt_8	int64	0.00	170
vol_2g_mb_6	float64	0.00	8406
vol_2g_mb_7	float64	0.00	8376
vol_2g_mb_8	float64	0.00	7622
vol_3g_mb_6	float64	0.00	7231
total_rech_amt_data_8	float64	0.00	1343

Max Recharge Data

Impute null values with 0 since null value means no recharge has happened.

telecom_high_val_cust['max_rech_data_6'] = telecom_high_val_cust['max_rech_data_6'].replace(np.NaN,0.0)
telecom_high_val_cust['max_rech_data_7'] = telecom_high_val_cust['max_rech_data_7'].replace(np.NaN,0.0)
telecom_high_val_cust['max_rech_data_8'] = telecom_high_val_cust['max_rech_data_8'].replace(np.NaN,0.0)

Columns with more than 40% missing data

All attributes that have more than 40% null values will be dropped.

# Calculate the percentage of missing values for each column
missing_percentages = telecom_high_val_cust.isnull().mean() * 100

# Identify columns with 40% or more missing values
columns_with_high_missing_values = missing_percentages[missing_percentages >= 40].index

# Dropping columns that have more than 40% missing values
telecom_high_val_cust = telecom_high_val_cust.drop(columns_with_high_missing_values,axis=1)

print("Dataframe shape after dropping columns with high missing values:",str(telecom_high_val_cust.shape))

Dataframe shape after dropping columns with high missing values: (20998, 131)

Minutes of Usage

Since mou columns have less than 5% missing data, we will impute them with 0 to indicate that this service was not used by the customer.

col_mou = [col for col in telecom_high_val_cust.columns if 'mou' in col]
telecom_high_val_cust[col_mou] = telecom_high_val_cust[col_mou].replace(np.NaN,0.0)

Inspect incoming and outgoing attributes for given months

• From the data analyzed through excel, it is identified that loc_ogmou* is the sum of all loc_og_mou columns. Likewise for std and total.
• The same is true for icmou* columns too.
• Therefore, these sum columns will be dropped for all months.

col_og_ic_mou = ['loc_og_mou_6','loc_og_mou_7','loc_og_mou_8',
                 'std_og_mou_6','std_og_mou_7','std_og_mou_8',
                 'total_og_mou_6','total_og_mou_7','total_og_mou_8',
                 'loc_ic_mou_6','loc_ic_mou_7','loc_ic_mou_8',
                 'std_ic_mou_6','std_ic_mou_7','std_ic_mou_8',
                 'total_ic_mou_6','total_ic_mou_7','total_ic_mou_8']

telecom_high_val_cust = telecom_high_val_cust.drop(col_og_ic_mou, axis=1)
print("Dataframe shape after dropping repetitive(correlated) columns:",str(telecom_high_val_cust.shape))

Dataframe shape after dropping repetitive(correlated) columns: (20998, 113)

Inspect other ic/og columns

Impute other ic/og columns with 0 since they have less than 5% mising data.

col_og_ic = [col for col in telecom_high_val_cust.columns if 'ic' in col or 'og' in col]
telecom_high_val_cust[col_og_ic] = telecom_high_val_cust[col_og_ic].replace(np.NaN,0.0)

STEP 2: Feature Engineering and Variable Transformation

Create new columns for average of 6th and 7th month attributes

telecom_df = telecom_high_val_cust.copy()

col_6_7 = [col for col in telecom_df.columns if '_6' in col or '_7' in col]
col_6_7_common = [col[:-2] for col in telecom_df.columns if '_6' in col or '_7' in col]
col_6_7_common = list(set(col_6_7_common))
for idx, col in enumerate(col_6_7_common):
    avg_col_6_7 = "avg_"+col+"_6_7"
    col_6 = col+"_6"
    col_7 = col+"_7"
    telecom_df[avg_col_6_7] = (telecom_df[col_6]  + telecom_df[col_7])/ 2

telecom_df = telecom_df.drop(col_6_7, axis = 1)
print("Dataframe shape after dropping transformed columns:",str(telecom_df.shape))

Dataframe shape after dropping transformed columns: (20998, 76)

Check correlation of variables for month 6_7

col_6_7 = [col for col in telecom_df.columns if '_6_7' in col]
plt.figure(figsize = (20,15))
sns.heatmap(telecom_df[col_6_7].corr(),annot = True)
plt.show()

We do not see any strong correlations that need to be addressed. One thing we notice is as recharge amount increases, average revenue per user also increases.

Final column list of independent variables or predictors

final_column_list = telecom_df.columns.to_list()
# Exclude target variable
final_column_list.remove('churn_probability')
print("Number of columns that will be used for model building:", len(final_column_list))

Number of columns that will be used for model building: 75

Outlier Treatment

# Determine skewness for all numeric variables
telecom_df.skew()

arpu_8                          18.145858
onnet_mou_8                      5.372370
offnet_mou_8                     4.758479
roam_ic_mou_8                   20.457859
roam_og_mou_8                   16.380833
loc_og_t2t_mou_8                15.260752
loc_og_t2m_mou_8                 4.097068
loc_og_t2f_mou_8                10.643422
loc_og_t2c_mou_8                16.890732
std_og_t2t_mou_8                 5.537740
std_og_t2m_mou_8                 6.833574
std_og_t2f_mou_8                15.496891
isd_og_mou_8                    81.027659
spl_og_mou_8                    15.237698
og_others_8                    107.578471
loc_ic_t2t_mou_8                11.065714
loc_ic_t2m_mou_8                 4.636332
loc_ic_t2f_mou_8                11.269502
std_ic_t2t_mou_8                20.611378
std_ic_t2m_mou_8                12.405555
std_ic_t2f_mou_8                36.590324
spl_ic_mou_8                     4.853747
isd_ic_mou_8                    17.759014
ic_others_8                     46.307587
total_rech_num_8                 2.800001
total_rech_amt_8                21.419988
max_rech_amt_8                   5.608865
last_day_rch_amt_8               7.054781
max_rech_data_8                  2.897118
vol_2g_mb_8                      8.138860
vol_3g_mb_8                      9.172511
monthly_2g_8                     2.890655
sachet_2g_8                      3.951159
monthly_3g_8                     5.010947
sachet_3g_8                     13.462701
aon                              0.986145
3g_vbc_8                         5.916601
churn_probability                3.032157
total_rech_amt_data_8           18.335753
avg_og_others_6_7               89.368833
avg_std_ic_t2m_mou_6_7          11.667077
avg_std_og_t2t_mou_6_7           4.412864
avg_total_rech_amt_data_6_7      9.434197
avg_std_ic_t2t_mou_6_7          24.908280
avg_3g_vbc_6_7                   4.800676
avg_spl_og_mou_6_7              20.505711
avg_ic_others_6_7               52.905578
avg_loc_ic_t2f_mou_6_7           9.501778
avg_total_rech_amt_6_7          21.257848
avg_std_ic_t2f_mou_6_7          33.144514
avg_monthly_2g_6_7               2.327500
avg_isd_ic_mou_6_7              31.879174
avg_loc_og_t2f_mou_6_7          10.783785
avg_roam_og_mou_6_7              9.009688
avg_spl_ic_mou_6_7              66.553173
avg_loc_og_t2c_mou_6_7          23.811981
avg_std_og_t2m_mou_6_7           4.237469
avg_isd_og_mou_6_7              81.977557
avg_sachet_2g_6_7                3.545139
avg_loc_og_t2m_mou_6_7           3.676190
avg_arpu_6_7                    20.881621
avg_last_day_rch_amt_6_7         5.431195
avg_total_rech_num_6_7           2.718843
avg_onnet_mou_6_7                3.898005
avg_vol_3g_mb_6_7                9.355957
avg_sachet_3g_6_7               11.164401
avg_std_og_t2f_mou_6_7          17.067401
avg_loc_ic_t2m_mou_6_7           4.356972
avg_roam_ic_mou_6_7             15.069792
avg_max_rech_amt_6_7             4.340042
avg_vol_2g_mb_6_7                5.958775
avg_max_rech_data_6_7            2.871534
avg_offnet_mou_6_7               3.156728
avg_loc_og_t2t_mou_6_7           9.560443
avg_monthly_3g_6_7               4.334295
avg_loc_ic_t2t_mou_6_7          10.305137
dtype: float64

# Visualize the percentage change between consecutive quantiles for the numeric columns in the telecom dataset. 
telecom_df.quantile(np.arange(0.9,1.01,0.01), axis=0, numeric_only=True).pct_change().mul(100).style.bar()

# Identify columns with outliers
# Using the percentage change between consecutive quantiles to identify columns with potentially large changes
# The threshold for considering a change as an outlier is set to a percentage change greater than 100

pct_change_99_1 = telecom_df.quantile(np.arange(0.9,1.01,0.01),axis=0, numeric_only=True).pct_change().mul(100).iloc[-1]
outlier_condition = pct_change_99_1 > 100
columns_with_outliers = pct_change_99_1[outlier_condition].index.values
print('Number of columns with outliers :', len(columns_with_outliers))

Number of columns with outliers : 74

# Treat the outliers - Values above the 99th percentile for each column are capped at the threshold
# Threshold = 99th percentile values

outlier_treatment = pd.DataFrame(columns=['Column', 'Outlier Threshold', 'Outliers replaced'])
for col in columns_with_outliers : 
    outlier_threshold = telecom_df[col].quantile(0.99)
    condition = telecom_df[col] > outlier_threshold
    df = pd.DataFrame({'Column' : col,'Outlier Threshold': outlier_threshold,'Outliers replaced': telecom_df.loc[condition,col].shape[0]}, index=[0])
    outlier_treatment = pd.concat([outlier_treatment, df])
    telecom_df.loc[condition, col] = outlier_threshold
outlier_treatment

	Column	Outlier Threshold	Outliers replaced
0	arpu_8	1971.592470	210
0	onnet_mou_8	2153.499100	210
0	offnet_mou_8	2157.510400	210
0	roam_ic_mou_8	274.131100	210
0	roam_og_mou_8	395.072000	210
0	loc_og_t2t_mou_8	899.539400	210
0	loc_og_t2m_mou_8	1065.090700	210
0	loc_og_t2f_mou_8	82.210900	210
0	loc_og_t2c_mou_8	27.212100	210
0	std_og_t2t_mou_8	1889.473300	210
0	std_og_t2m_mou_8	1873.920100	210
0	std_og_t2f_mou_8	34.980600	210
0	isd_og_mou_8	28.797200	210
0	spl_og_mou_8	71.797500	210
0	og_others_8	0.000000	126
0	loc_ic_t2t_mou_8	619.390100	210
0	loc_ic_t2m_mou_8	955.963900	210
0	loc_ic_t2f_mou_8	179.601900	210
0	std_ic_t2t_mou_8	211.139000	210
0	std_ic_t2m_mou_8	357.242300	210
0	std_ic_t2f_mou_8	45.050200	210
0	spl_ic_mou_8	0.560000	206
0	isd_ic_mou_8	227.415100	210
0	ic_others_8	21.350900	210
0	total_rech_num_8	44.000000	207
0	total_rech_amt_8	2328.120000	210
0	max_rech_amt_8	951.000000	199
0	last_day_rch_amt_8	565.000000	207
0	max_rech_data_8	455.000000	120
0	vol_2g_mb_8	1449.599000	210
0	vol_3g_mb_8	3917.502400	210
0	monthly_2g_8	2.000000	27
0	sachet_2g_8	12.000000	163
0	monthly_3g_8	3.000000	102
0	sachet_3g_8	5.000000	149
0	3g_vbc_8	1965.210600	210
0	total_rech_amt_data_8	5688.720000	210
0	avg_og_others_6_7	4.595000	209
0	avg_std_ic_t2m_mou_6_7	373.329800	210
0	avg_std_og_t2t_mou_6_7	1788.015650	210
0	avg_total_rech_amt_data_6_7	5134.680000	210
0	avg_std_ic_t2t_mou_6_7	196.187250	210
0	avg_3g_vbc_6_7	1817.028400	210
0	avg_spl_og_mou_6_7	63.476800	210
0	avg_ic_others_6_7	18.902100	210
0	avg_loc_ic_t2f_mou_6_7	190.901200	210
0	avg_total_rech_amt_6_7	2202.090000	210
0	avg_std_ic_t2f_mou_6_7	46.967250	210
0	avg_monthly_2g_6_7	1.500000	80
0	avg_isd_ic_mou_6_7	222.025650	210
0	avg_loc_og_t2f_mou_6_7	84.923350	210
0	avg_roam_og_mou_6_7	448.911150	210
0	avg_spl_ic_mou_6_7	0.410000	209
0	avg_loc_og_t2c_mou_6_7	21.565150	210
0	avg_std_og_t2m_mou_6_7	1859.086850	210
0	avg_isd_og_mou_6_7	39.886150	210
0	avg_sachet_2g_6_7	11.000000	208
0	avg_loc_og_t2m_mou_6_7	1071.498200	210
0	avg_arpu_6_7	1841.591235	210
0	avg_last_day_rch_amt_6_7	549.500000	208
0	avg_total_rech_num_6_7	45.000000	206
0	avg_onnet_mou_6_7	2059.418800	210
0	avg_vol_3g_mb_6_7	3482.782950	210
0	avg_sachet_3g_6_7	4.000000	192
0	avg_std_og_t2f_mou_6_7	36.938150	210
0	avg_loc_ic_t2m_mou_6_7	929.915900	210
0	avg_roam_ic_mou_6_7	272.755800	210
0	avg_max_rech_amt_6_7	790.000000	209
0	avg_vol_2g_mb_6_7	1364.016450	210
0	avg_max_rech_data_6_7	449.000000	196
0	avg_offnet_mou_6_7	2157.215050	210
0	avg_loc_og_t2t_mou_6_7	975.177750	210
0	avg_monthly_3g_6_7	2.500000	164
0	avg_loc_ic_t2t_mou_6_7	631.440950	210

STEP 3: Exploratory Data Analysis

# Copy the dataset and perform eda so that existing columns are not disturbed
telecom_eda = telecom_df.copy()

Separate churn and non-churn customers

churn = telecom_eda[telecom_eda['churn_probability']==1]
not_churn = telecom_eda[telecom_eda['churn_probability']==0]
print("Churn dataset:",churn.shape)
print("Non-churn dataset:",not_churn.shape)

Churn dataset: (1735, 76)
Non-churn dataset: (19263, 76)

Customer Churn Probability as Pie Chart

# Create labels for better graph understanding
telecom_eda['churn_labels'] = telecom_eda['churn_probability'].map({0: 'No', 1: 'Yes'})
lab = telecom_eda['churn_labels'].value_counts().keys().tolist()
val = telecom_eda['churn_labels'].value_counts().values.tolist()

trace = go.Pie(labels = lab,
               values = val,
               marker = dict(colors = ['royalblue', 'lime'],
                             line = dict(color = "white",
                                         width = 1.3)
                            ),
              rotation = 90,
              hoverinfo = "label+value+text",
              hole = .5
              )
layout = go.Layout(dict(title = "Customer Churn in Data",
                        plot_bgcolor = "rgb(243,243,243)",
                        paper_bgcolor = "rgb(243,243,243)",
                       )
                  )
data = [trace]
fig = go.Figure(data = data, layout = layout)
pyo.iplot(fig)

We see that 8.26% of the high-value customers currently churn or move to other networks. Our goal in this case study is to:

• predict if a customer will churn or not
• what are the factors that contribute to churning of customers

Histogram to view distribution of customers churning across various attributes

def histogram(column):
    trace1 = go.Histogram(x = not_churn[column],
                         histnorm="percent",
                         name="Non-churn Customers",
                         marker = dict(line = dict(width = 0.5,
                                                  color = "black"
                                                  )
                                      ),
                         opacity = .9
                         )
    trace2 = go.Histogram(x = churn[column],
                         histnorm="percent",
                         name="Churn Customers",
                         marker = dict(line = dict(width = 0.5,
                                                  color = "black"
                                                  )
                                      ),
                         opacity = .9
                         )
    data = [trace1, trace2]
    layout = go.Layout(dict(title = column + " - Distribution in Customer Churn",
                           plot_bgcolor = "rgb(243,243,243)",
                           paper_bgcolor = "rgb(243,243,243)",
                           xaxis = dict(gridcolor = 'rgb(255, 255, 255)',
                                       title = column,
                                       zerolinewidth = 1,
                                       ticklen = 5,
                                       gridwidth = 2
                                       ),
                           yaxis = dict(gridcolor = 'rgb(255, 255, 255)',
                                       title = column,
                                       zerolinewidth = 1,
                                       ticklen = 5,
                                       gridwidth = 2
                                       ),
                           )
                      )
    fig = go.Figure(data=data, layout=layout)
    pyo.iplot(fig)
    
for i in final_column_list:
    histogram(i)

Observation:

• It can be observed that customers churn when their charges are close to 0. It seems that when they are less liable to the telecom operator, they churn quickly.
• One recommendation would be to try and retain customers and make them buy larger packs that cost high so that they do not leave the telecom operator.

Customer churn view by tenure

# Create new column tenure from age on network (aon) for better analysis

# Age on network(aon) - number of days the customer is using the operator T network
telecom_eda['tenure'] = telecom_eda['aon']/30

tenure_range = [0, 6, 12, 24, 60, 61]
tenure_label = [ '0-6 Months', '6-12 Months', '1-2 Yrs', '2-5 Yrs', '5 Yrs and above']
telecom_eda['tenure_range'] = pd.cut(telecom_eda['tenure'], tenure_range, labels=tenure_label)
telecom_eda['tenure_range'].head()

4     1-2 Yrs
15    1-2 Yrs
23    1-2 Yrs
24    2-5 Yrs
27    1-2 Yrs
Name: tenure_range, dtype: category
Categories (5, object): ['0-6 Months' < '6-12 Months' < '1-2 Yrs' < '2-5 Yrs' < '5 Yrs and above']

churn = telecom_eda[telecom_eda['churn_probability']==1]
not_churn = telecom_eda[telecom_eda['churn_probability']==0]

# Plot Graph to Display Percentage of Customers Churned Across Various Tenures
tenure_total = telecom_eda['tenure_range'].value_counts().sort_index()
tenure_churn = churn['tenure_range'].value_counts().sort_index()
tenure_churn_per = (tenure_churn.values / tenure_total.values)*100
tenure_nchurn = not_churn['tenure_range'].value_counts().sort_index()
tenure_nchurn_per = (tenure_nchurn.values / tenure_total.values)*100

fig = go.Figure()
fig.add_trace(go.Bar(
    x=tenure_total.index,
    y=tenure_churn_per,
    name='Customer Churning',
    offset=-0.2,
    marker=dict(color='Red')
))
fig.add_trace(go.Bar(
    x=tenure_total.index,
    y=tenure_nchurn_per,
    name='Customer Not Churning',
    offset=+0.2,
    marker=dict(color='Blue')
))
bar_width = 0.4
fig.update_traces(width=bar_width)
fig.update_layout(
    title='Customer Churning Status Per Tenure Range',
    xaxis_title='Tenure Range',
    yaxis_title='Percentage of Customers',
)
fig.show()

Observation:

• Customers who stay longer with the telecom operator tend to churn less.
• Schemes can be brought in to attract new customers so thatthey stay longer with the network.

# scatter plot for total charges and max recharges by tenure group

def plot_tenure_scatter(tenure_range, color) :
    tracer = go.Scatter(x = telecom_eda[telecom_eda["tenure_range"] == tenure_range]["total_rech_amt_8"],
                        y = telecom_eda[telecom_eda["tenure_range"] == tenure_range]["max_rech_amt_8"],
                        mode = "markers",
                        marker = dict(line = dict(color = "black",
                                                  width = .2),
                                      size = 4, color = color,
                                      symbol = "diamond-dot",
                                    ),
                        name = tenure_range,
                        opacity = .9
                        )
    return tracer
    
# scatter plot for total charges and max recharges by churn group
def plot_churn_scatter(churn, color) :
    tracer = go.Scatter(x = telecom_eda[telecom_eda["churn_labels"] == churn]["total_rech_amt_8"],
                        y = telecom_eda[telecom_eda["churn_labels"] == churn]["max_rech_amt_8"],
                        mode = "markers",
                        marker = dict(line = dict(color = "black",
                                                  width = .2),
                                      size = 4, color = color,
                                      symbol = "diamond-dot",
                                    ),
                        name = churn,
                        opacity = .9
                        )
    return tracer

trace1 = plot_tenure_scatter("0-6 Months","#FF3366")
trace2 = plot_tenure_scatter("6-12 Months","#6666FF")
trace3 = plot_tenure_scatter("1-2 Yrs","#00FF00")
trace4 = plot_tenure_scatter("2-5 Yrs","#996600")
trace5 = plot_tenure_scatter("5 Yrs and above","grey")
trace6 = plot_churn_scatter("Yes","red")
trace7 = plot_churn_scatter("No","blue")

data1 = [trace1, trace3, trace4, trace5, trace2]
data2 = [trace7, trace6]

#layout
def layout_title(title):
    layout = go.Layout(dict(title = title,
                           plot_bgcolor = "rgb(243,243,243)",
                           paper_bgcolor = "rgb(243,243,243)",
                           xaxis = dict(gridcolor = 'rgb(255,255,255)',
                                        title = "Total charges",
                                       zerolinewidth=1, ticklen = 5, gridwidth=2),
                           yaxis = dict(gridcolor = 'rgb(255,255,255)',
                                        title = "Maximum charges",
                                       zerolinewidth=1, ticklen = 5, gridwidth=2),
                           height = 600
                           )
                )
    return layout

layout1 = layout_title("Monthly Charges & Max Charges by Tenure Range for August")
layout2 = layout_title("Monthly Charges & Max Charges by Churn Probability for August ")
fig1 = go.Figure(data = data1, layout = layout1)
fig2 = go.Figure(data = data2, layout = layout2)
pyo.iplot(fig1)
pyo.iplot(fig2)

Observation:

• Comparing the above scatter plots it is clear that when total charges and max charges are less, customers churn more.
• Customers who have been with the operator for less than a year tend to churn more as can be clearly seen as an overlaping pattern from graph 2.

Scatter plots to understand attribute spread across months Vs churn

def scatter_month(x_data,y_data):
    fig = px.scatter(telecom_eda, x=x_data, y=y_data, color='churn_labels')
    pyo.iplot(fig)

scatter_month('avg_total_rech_amt_6_7', 'total_rech_amt_8')

scatter_month('avg_max_rech_data_6_7', 'max_rech_data_8')

scatter_month('avg_last_day_rch_amt_6_7', 'last_day_rch_amt_8')

scatter_month('avg_arpu_6_7','arpu_8')

scatter_month('avg_roam_og_mou_6_7','roam_og_mou_8')

scatter_month('avg_roam_ic_mou_6_7','roam_ic_mou_8')

Observations:

• Above scatter plots are a few plots that strongly reiterate that if a customer suddenly drops usage in the last month, he will most likely churn.
• It is very clear that the red spots that indicate churning are concentrated where month_8 usage has been close to 0.
• Attributes pertaining to month_8 will be very crucial in deiciding if a customer will churn.

STEP 4: Model Building

Train-test split

# Independent variables/Features in X
X = telecom_df.drop('churn_probability', axis = 1)

# Target variable in y
y = telecom_df['churn_probability']

# Splitting data into train and test set 80:20
X_train, X_test, y_train, y_test = train_test_split(X, y, train_size=0.8, test_size=0.2, random_state=100)

Imbalance Correction

We know there is data imbalance in the dataset as seen in the pie chart in EDA. 91.7% customers do not churn, only 8.26% customers churn. We will deal with the imbalance using SMOTE (Synthetic Minority Oversampling Technique)

# SMOTE
sm = SMOTE(random_state=42)
X_train, y_train = sm.fit_resample(X_train, y_train)

Scale the data

As seen in the metadat matrix of the dataset, the values of attributes lie in various ranges. Therefore, we will scale them using standard scaler technique.

# Instantiate the Scaler
scaler = StandardScaler()

# Fit the data into scaler and transform
X_train[final_column_list] = scaler.fit_transform(X_train[final_column_list])

# Transform the test set
X_test[final_column_list] = scaler.transform(X_test[final_column_list])

Modeling with PCA

• PCA (Principal Component Analysis) is a statistical technique used to simplify the complexity in high-dimensional data while retaining trends and patterns.
• PCA transforms the original variables into a new set of uncorrelated variables called principal components.
• PCA helps in dimensionality reduction, noise reduction, and visualization of data.

# Instantiate PCA
pca = PCA(random_state=42)

# Fit train set on PCA
pca.fit(X_train)

# Principal components
pca.components_

# Cumuliative variance of the principal components
variance_cumu = np.cumsum(pca.explained_variance_ratio_)

# Plotting scree plot
fig = plt.figure(figsize=[12,8])
plt.vlines(x=46, ymax=1, ymin=0, colors="r", linestyles="--")
plt.hlines(y=0.95, xmax=80, xmin=0, colors="g", linestyles="--")
plt.plot(variance_cumu)
plt.xlabel('Number of Components')
plt.ylabel("Cumulative variance explained")
plt.show()

About 45 components explain roughly about 95% of the variance in the dataset. Therefore, we will build PCA with 45 components.

# Instantiate PCA with 45 components
pca_final = IncrementalPCA(n_components=45)

# Fit and transform the X_train
X_train_pca = pca_final.fit_transform(X_train)
X_test_pca = pca_final.transform(X_test)

def model_metrics(y_data, y_data_pred):
    table_data = []
    confusion = metrics.confusion_matrix(y_data, y_data_pred)

    TP = confusion[1,1] # true positive 
    TN = confusion[0,0] # true negatives
    FP = confusion[0,1] # false positives
    FN = confusion[1,0] # false negatives

    # Accuracy
    accuracy = round(metrics.accuracy_score(y_data, y_data_pred),3)
    

    # Sensitivity
    sensitivity = round(TP/float(FN + TP),3)

    # Specificity
    specificity = round(TN / float(TN+FP),3)
    
    #Precision
    precision = round(TP/float(TP + FP),3)
    
    #F1 score
    f1_score = round(2*precision*sensitivity/(precision + sensitivity),3)
    
    # Add metrics to table data
    table_data.append(["Confusion matrix", confusion])
    table_data.append(["Accuracy", accuracy])
    table_data.append(["Sensitivity/Recall", sensitivity])
    table_data.append(["Specificity", specificity])
    table_data.append(["Precision", precision])
    table_data.append(["F1-score", f1_score])

    # Print the formatted table
    print(tabulate(table_data, headers=["Metric", "Value"], tablefmt="grid"))

Model 1 - Logistic Regression with PCA

# Creating KFold object with 5 splits
folds = StratifiedKFold(n_splits=3, shuffle=True, random_state=4)

# Specify params
params = {"C": [0.01, 0.1, 1, 10, 100, 1000]
         }

# Specifing score as recall as we are more focused on acheiving the higher sensitivity than the accuracy
log1_model_cv = GridSearchCV(estimator = LogisticRegression(),
                        param_grid = params, 
                        scoring= 'roc_auc', 
                        cv = folds, 
                        verbose = 1,
                        return_train_score=True) 

# Fit the model
log1_model_cv.fit(X_train_pca, y_train)

Fitting 3 folds for each of 6 candidates, totalling 18 fits

GridSearchCV(cv=StratifiedKFold(n_splits=3, random_state=4, shuffle=True),
             estimator=LogisticRegression(),
             param_grid={'C': [0.01, 0.1, 1, 10, 100, 1000]},
             return_train_score=True, scoring='roc_auc', verbose=1)

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

# results of grid search CV
cv_results = pd.DataFrame(log1_model_cv.cv_results_)
# print best hyperparameters
print("Best AUC: ", log1_model_cv.best_score_)
print("Best hyperparameters: ", log1_model_cv.best_params_)

Best AUC:  0.914145455747912
Best hyperparameters:  {'C': 100}

# Instantiate the model with best C

logistic_pca = LogisticRegression(C=100)

# Fit the model on the train set
log1_pca_model = logistic_pca.fit(X_train_pca, y_train)

# Predictions on the train set
y_train_pred = log1_pca_model.predict(X_train_pca)

print("\nTraining set:")
model_metrics(y_train, y_train_pred)

# Prediction on the test set
y_test_pred = log1_pca_model.predict(X_test_pca)

print("\nValidation set:")
model_metrics(y_test, y_test_pred)

Training set:
+--------------------+-----------------+
| Metric             | Value           |
+====================+=================+
| Confusion matrix   | [[12661  2739]  |
|                    |  [ 2187 13213]] |
+--------------------+-----------------+
| Accuracy           | 0.84            |
+--------------------+-----------------+
| Sensitivity/Recall | 0.858           |
+--------------------+-----------------+
| Specificity        | 0.822           |
+--------------------+-----------------+
| Precision          | 0.828           |
+--------------------+-----------------+
| F1-score           | 0.843           |
+--------------------+-----------------+

Validation set:
+--------------------+---------------+
| Metric             | Value         |
+====================+===============+
| Confusion matrix   | [[3171  692]  |
|                    |  [  63  274]] |
+--------------------+---------------+
| Accuracy           | 0.82          |
+--------------------+---------------+
| Sensitivity/Recall | 0.813         |
+--------------------+---------------+
| Specificity        | 0.821         |
+--------------------+---------------+
| Precision          | 0.284         |
+--------------------+---------------+
| F1-score           | 0.421         |
+--------------------+---------------+

Model 2 - SVM with PCA

# SVM Linear Vs RBF - to determine which kernel to be used

model = SVC(kernel="linear")
model.fit(X_train_pca, y_train)
y_test_pred = model.predict(X_test_pca)
accuracy = round(metrics.accuracy_score(y_test, y_test_pred),3)
print("Accuracy of linear model is:", accuracy)

Accuracy of linear model is: 0.832

model = SVC(kernel="rbf")
model.fit(X_train_pca, y_train)
y_test_pred = model.predict(X_test_pca)
accuracy = round(metrics.accuracy_score(y_test, y_test_pred),3)
print("Accuracy of rbf model is:", accuracy)

Accuracy of rbf model is: 0.885

We see that SVM with Radial Basis Function gives better accuracy. Therefore, let's tune parameters for rbf model.

# specify range of hyperparameters

folds = StratifiedKFold(n_splits = 3, shuffle = True, random_state = 100)
hyper_params = [ {'gamma': [1e-1, 1e-2],
                     'C': [1, 10]}]

# specify model with RBF kernel
model = SVC(kernel="rbf")

# set up GridSearchCV()
svm1_model_cv = GridSearchCV(estimator = model, 
                        param_grid = hyper_params, 
                        scoring= 'accuracy', 
                        cv = folds, 
                        verbose = 1,
                        refit = True,
                        return_train_score=True,
                        n_jobs = -1)      

# fit the model
svm1_model_cv.fit(X_train_pca, y_train) 

Fitting 3 folds for each of 4 candidates, totalling 12 fits

GridSearchCV(cv=StratifiedKFold(n_splits=3, random_state=100, shuffle=True),
             estimator=SVC(), n_jobs=-1,
             param_grid=[{'C': [1, 10], 'gamma': [0.1, 0.01]}],
             return_train_score=True, scoring='accuracy', verbose=1)

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

# cv results
cv_results = pd.DataFrame(svm1_model_cv.cv_results_)
cv_results

	mean_fit_time	std_fit_time	mean_score_time	std_score_time	param_C	param_gamma	params	split0_test_score	split1_test_score	split2_test_score	mean_test_score	std_test_score	rank_test_score	split0_train_score	split1_train_score	split2_train_score	mean_train_score	std_train_score
0	176.286576	12.767550	25.489950	1.543002	1	0.1	{'C': 1, 'gamma': 0.1}	0.959482	0.960748	0.955776	0.958669	0.002110	2	0.982078	0.982321	0.983150	0.982516	0.000459
1	35.476198	0.300432	25.920587	0.431521	1	0.01	{'C': 1, 'gamma': 0.01}	0.907178	0.910393	0.905903	0.907825	0.001889	4	0.917791	0.916086	0.916967	0.916948	0.000696
2	168.488312	22.244605	24.981936	1.646983	10	0.1	{'C': 10, 'gamma': 0.1}	0.972144	0.971267	0.967076	0.970162	0.002212	1	0.997273	0.996786	0.997175	0.997078	0.000210
3	35.726973	0.368254	19.378474	0.277905	10	0.01	{'C': 10, 'gamma': 0.01}	0.937957	0.937275	0.936295	0.937175	0.000682	3	0.952905	0.953197	0.952713	0.952938	0.000199

# print best hyperparameters
print("Best AUC: ", svm1_model_cv.best_score_)
print("Best hyperparameters: ", svm1_model_cv.best_params_)

Best AUC:  0.9701622374528146
Best hyperparameters:  {'C': 10, 'gamma': 0.1}

# Building the model with optimal hyperparameters
svm1_pca_model = SVC(C=10, gamma=0.1, kernel="rbf")

svm1_pca_model.fit(X_train_pca, y_train)

# Predictions on the train set
y_train_pred = svm1_pca_model.predict(X_train_pca)

print("\nTraining set:")
model_metrics(y_train, y_train_pred)

# Prediction on the test set
y_test_pred = svm1_pca_model.predict(X_test_pca)

print("\nValidation set:")
model_metrics(y_test, y_test_pred)

Training set:
+--------------------+-----------------+
| Metric             | Value           |
+====================+=================+
| Confusion matrix   | [[15310    90]  |
|                    |  [    2 15398]] |
+--------------------+-----------------+
| Accuracy           | 0.997           |
+--------------------+-----------------+
| Sensitivity/Recall | 1.0             |
+--------------------+-----------------+
| Specificity        | 0.994           |
+--------------------+-----------------+
| Precision          | 0.994           |
+--------------------+-----------------+
| F1-score           | 0.997           |
+--------------------+-----------------+

Validation set:
+--------------------+---------------+
| Metric             | Value         |
+====================+===============+
| Confusion matrix   | [[3742  121]  |
|                    |  [ 210  127]] |
+--------------------+---------------+
| Accuracy           | 0.921         |
+--------------------+---------------+
| Sensitivity/Recall | 0.377         |
+--------------------+---------------+
| Specificity        | 0.969         |
+--------------------+---------------+
| Precision          | 0.512         |
+--------------------+---------------+
| F1-score           | 0.434         |
+--------------------+---------------+

# specify range of hyperparameters

folds = StratifiedKFold(n_splits = 3, shuffle = True, random_state = 100)
hyper_params = [ {'gamma': [1e-1, 1e-2],
                     'C': [1, 10],
                  'class_weight': ['balanced']}]

# specify model with RBF kernel
model = SVC(kernel="rbf")

# set up GridSearchCV()
svm2_model_cv = GridSearchCV(estimator = model, 
                        param_grid = hyper_params, 
                        scoring= 'accuracy', 
                        cv = folds, 
                        verbose = 1,
                        refit = True,
                        return_train_score=True,
                        n_jobs = -1)      

# fit the model
svm2_model_cv.fit(X_train_pca, y_train) 

Fitting 3 folds for each of 4 candidates, totalling 12 fits

GridSearchCV(cv=StratifiedKFold(n_splits=3, random_state=100, shuffle=True),
             estimator=SVC(), n_jobs=-1,
             param_grid=[{'C': [1, 10], 'class_weight': ['balanced'],
                          'gamma': [0.1, 0.01]}],
             return_train_score=True, scoring='accuracy', verbose=1)

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

# print best hyperparameters
print("Best AUC: ", svm2_model_cv.best_score_)
print("Best hyperparameters: ", svm2_model_cv.best_params_)

Best AUC:  0.9701622374528146
Best hyperparameters:  {'C': 10, 'class_weight': 'balanced', 'gamma': 0.1}

# Building the model with optimal hyperparameters
svm2_pca_model = SVC(C=10, class_weight="balanced", gamma=0.1, kernel="rbf", degree=2)

svm2_pca_model.fit(X_train_pca, y_train)

# Predictions on the train set
y_train_pred = svm2_pca_model.predict(X_train_pca)

print("\nTraining set:")
model_metrics(y_train, y_train_pred)

# Prediction on the test set
y_test_pred = svm2_pca_model.predict(X_test_pca)

print("\nValidation set:")
model_metrics(y_test, y_test_pred)

Training set:
+--------------------+-----------------+
| Metric             | Value           |
+====================+=================+
| Confusion matrix   | [[15310    90]  |
|                    |  [    2 15398]] |
+--------------------+-----------------+
| Accuracy           | 0.997           |
+--------------------+-----------------+
| Sensitivity/Recall | 1.0             |
+--------------------+-----------------+
| Specificity        | 0.994           |
+--------------------+-----------------+
| Precision          | 0.994           |
+--------------------+-----------------+
| F1-score           | 0.997           |
+--------------------+-----------------+

Validation set:
+--------------------+---------------+
| Metric             | Value         |
+====================+===============+
| Confusion matrix   | [[3742  121]  |
|                    |  [ 210  127]] |
+--------------------+---------------+
| Accuracy           | 0.921         |
+--------------------+---------------+
| Sensitivity/Recall | 0.377         |
+--------------------+---------------+
| Specificity        | 0.969         |
+--------------------+---------------+
| Precision          | 0.512         |
+--------------------+---------------+
| F1-score           | 0.434         |
+--------------------+---------------+

Model 3 - Random Forest with PCA

# Basic RFC to check accuracy

rfc = RandomForestClassifier()
rfc.fit(X_train_pca,y_train)
y_test_pred = rfc.predict(X_test_pca)
accuracy = round(metrics.accuracy_score(y_test, y_test_pred),3)
print("Accuracy of basic rfc model is:", accuracy)

Accuracy of basic rfc model is: 0.913

We can see that Random Forest Classifier performs very well and therefore, we will go ahead and tune its parameters.

# define a function for tuning parameters for Random Forest individually
def hpt(params, x_train, y_train, scoring, max_depth=0):
    folds = StratifiedKFold(n_splits = 3, shuffle = True, random_state = 100)
    if(max_depth==0):
        rf = RandomForestClassifier()
    else :
        rf = RandomForestClassifier(max_depth=max_depth)
    
    grid_search = GridSearchCV(estimator = rf, 
                           param_grid = params, 
                           cv = folds,
                           n_jobs = -1,
                           scoring = scoring,
                           refit = True,
                           return_train_score=True)
    
    grid_search.fit(x_train, y_train)
    scores = grid_search.cv_results_

    for key in params.keys():
        hyperparameters = key
        break

    # plotting accuracies for parameters
    plt.figure(figsize=(16,5))
    plt.plot(scores["param_"+hyperparameters], scores["mean_train_score"], label="training accuracy")
    plt.plot(scores["param_"+hyperparameters], scores["mean_test_score"], label="test accuracy")
    plt.xlabel(hyperparameters)
    plt.ylabel("Accuracy")
    plt.legend()
    plt.show()

# Tuning max_depth
params = {'max_depth': range(2, 40, 5)}
hpt(params,X_train_pca,y_train, "accuracy")

We see that as max_depth incraeses, the training and test accuracy increases. We will use 12 and 17 for gridview search as they have peaks.

# Tuning min_samples_leaf
params = {'min_samples_leaf': range(1, 100, 10)}
hpt(params,X_train_pca,y_train,"accuracy")

We will use values between 10 and 20 to train our model as there is a steep dip after 20.

# Tuning min_samples_split
params = {'min_samples_split': range(10, 100, 10)}
hpt(params,X_train_pca,y_train,"accuracy")

We will use values between 10 and 20 as the first dip is at 20.

# Tuning n_estimators
params = {'n_estimators': range(50, 1000, 100)}
hpt(params,X_train_pca,y_train,"accuracy")

We will use 150 since the test accuracy is slightly dipping post that.

# Tuning max_features
params = {'max_features': range(10,50,10)}
hpt(params,X_train_pca,y_train,"accuracy",5)

We can use 15 or log2 for max features

folds = StratifiedKFold(n_splits = 3, shuffle = True, random_state = 100)
params = {
    'max_depth': [12,17],
    'min_samples_leaf': [10,20],
    'min_samples_split': [10,20],
    'n_estimators': [150], 
    'max_features': [15]
}

# Create a based model
rfc1 = RandomForestClassifier(random_state=42, n_jobs=-1)
# Instantiate the grid search model
grid_search = GridSearchCV(estimator = rfc1, 
                           param_grid = params, 
                           cv = folds,
                           n_jobs = -1,
                           verbose = 1, 
                           scoring = "accuracy",
                           refit = True,
                           return_train_score=True)

# Fit the model
grid_search.fit(X_train_pca, y_train)

Fitting 3 folds for each of 8 candidates, totalling 24 fits

GridSearchCV(cv=StratifiedKFold(n_splits=3, random_state=100, shuffle=True),
             estimator=RandomForestClassifier(n_jobs=-1, random_state=42),
             n_jobs=-1,
             param_grid={'max_depth': [12, 17], 'max_features': [15],
                         'min_samples_leaf': [10, 20],
                         'min_samples_split': [10, 20], 'n_estimators': [150]},
             return_train_score=True, scoring='accuracy', verbose=1)

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

# print best hyperparameters
print("Best AUC: ", grid_search.best_score_)
print("Best hyperparameters: ", grid_search.best_params_)

Best AUC:  0.9119480580874569
Best hyperparameters:  {'max_depth': 17, 'max_features': 15, 'min_samples_leaf': 10, 'min_samples_split': 10, 'n_estimators': 150}

# Building the model with optimal hyperparameters

## check these values
rf1_pca_model = RandomForestClassifier(bootstrap=True,
                             max_depth=17,
                             min_samples_leaf=10, 
                             min_samples_split=10,
                             max_features=15,
                             n_estimators=150)

rf1_pca_model.fit(X_train_pca, y_train)

# Predictions on the train set
y_train_pred = rf1_pca_model.predict(X_train_pca)

print("\nTraining set:")
model_metrics(y_train, y_train_pred)

# Prediction on the test set
y_test_pred = rf1_pca_model.predict(X_test_pca)

print("\nValidation set:")
model_metrics(y_test, y_test_pred)

Training set:
+--------------------+-----------------+
| Metric             | Value           |
+====================+=================+
| Confusion matrix   | [[14625   775]  |
|                    |  [  348 15052]] |
+--------------------+-----------------+
| Accuracy           | 0.964           |
+--------------------+-----------------+
| Sensitivity/Recall | 0.977           |
+--------------------+-----------------+
| Specificity        | 0.95            |
+--------------------+-----------------+
| Precision          | 0.951           |
+--------------------+-----------------+
| F1-score           | 0.964           |
+--------------------+-----------------+

Validation set:
+--------------------+---------------+
| Metric             | Value         |
+====================+===============+
| Confusion matrix   | [[3556  307]  |
|                    |  [ 111  226]] |
+--------------------+---------------+
| Accuracy           | 0.9           |
+--------------------+---------------+
| Sensitivity/Recall | 0.671         |
+--------------------+---------------+
| Specificity        | 0.921         |
+--------------------+---------------+
| Precision          | 0.424         |
+--------------------+---------------+
| F1-score           | 0.52          |
+--------------------+---------------+

# Building the model with optimal hyperparameters

## check these values
rf3_pca_model = RandomForestClassifier(bootstrap=True,
                             min_samples_split=10,
                             max_features='log2',
                             #criterion='entropy',
                             class_weight='balanced',
                             n_estimators=200)

rf3_pca_model.fit(X_train_pca, y_train)

# Predictions on the train set
y_train_pred = rf3_pca_model.predict(X_train_pca)

print("\nTraining set:")
model_metrics(y_train, y_train_pred)

# Prediction on the test set
y_test_pred = rf3_pca_model.predict(X_test_pca)

print("\nValidation set:")
model_metrics(y_test, y_test_pred)

Training set:
+--------------------+-----------------+
| Metric             | Value           |
+====================+=================+
| Confusion matrix   | [[15269   131]  |
|                    |  [   28 15372]] |
+--------------------+-----------------+
| Accuracy           | 0.995           |
+--------------------+-----------------+
| Sensitivity/Recall | 0.998           |
+--------------------+-----------------+
| Specificity        | 0.991           |
+--------------------+-----------------+
| Precision          | 0.992           |
+--------------------+-----------------+
| F1-score           | 0.995           |
+--------------------+-----------------+

Validation set:
+--------------------+---------------+
| Metric             | Value         |
+====================+===============+
| Confusion matrix   | [[3639  224]  |
|                    |  [ 145  192]] |
+--------------------+---------------+
| Accuracy           | 0.912         |
+--------------------+---------------+
| Sensitivity/Recall | 0.57          |
+--------------------+---------------+
| Specificity        | 0.942         |
+--------------------+---------------+
| Precision          | 0.462         |
+--------------------+---------------+
| F1-score           | 0.51          |
+--------------------+---------------+

Model 4 - XGBoost with PCA

# Basic XGBoost to check accuracy
xgb = XGBClassifier()
xgb.fit(X_train_pca, y_train)
y_test_pred = xgb.predict(X_test_pca)
accuracy = round(metrics.accuracy_score(y_test, y_test_pred),3)
print("Accuracy of basic XGB model is:", accuracy)

Accuracy of basic XGB model is: 0.905

Basic accuracy of XGB is very good. Therefore, let's tune the parameters for it.

folds = StratifiedKFold(n_splits = 3, shuffle = True, random_state = 100)
params = {
    'learning_rate': [0.1,0.2,0.3],
    'max_depth': [2,5], 
    'n_estimators': [200],
    'subsample':[0.3,0.4,0.5,1.0]
}

# Create a xbg model
xgb1 = XGBClassifier(random_state=42, n_jobs=-1)

# Instantiate the grid search model
grid_search = GridSearchCV(estimator = xgb1, 
                           param_grid = params, 
                           cv = folds,
                           n_jobs = -1,
                           verbose = 1, 
                           scoring = "accuracy",
                           refit = True,
                           return_train_score=True)

# Fit the model
grid_search.fit(X_train_pca, y_train)

Fitting 3 folds for each of 24 candidates, totalling 72 fits

GridSearchCV(cv=StratifiedKFold(n_splits=3, random_state=100, shuffle=True),
             estimator=XGBClassifier(base_score=None, booster=None,
                                     callbacks=None, colsample_bylevel=None,
                                     colsample_bynode=None,
                                     colsample_bytree=None, device=None,
                                     early_stopping_rounds=None,
                                     enable_categorical=False, eval_metric=None,
                                     feature_types=None, gamma=None,
                                     grow_policy=None, importance_ty...
                                     max_delta_step=None, max_depth=None,
                                     max_leaves=None, min_child_weight=None,
                                     missing=nan, monotone_constraints=None,
                                     multi_strategy=None, n_estimators=None,
                                     n_jobs=-1, num_parallel_tree=None,
                                     random_state=42, ...),
             n_jobs=-1,
             param_grid={'learning_rate': [0.1, 0.2, 0.3], 'max_depth': [2, 5],
                         'n_estimators': [200],
                         'subsample': [0.3, 0.4, 0.5, 1.0]},
             return_train_score=True, scoring='accuracy', verbose=1)

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

# print best hyperparameters
print("Best AUC: ", grid_search.best_score_)
print("Best hyperparameters: ", grid_search.best_params_)

Best AUC:  0.9449349868100266
Best hyperparameters:  {'learning_rate': 0.3, 'max_depth': 5, 'n_estimators': 200, 'subsample': 1.0}

# Building the model with optimal hyperparameters

## check these values
xgb1_pca_model = XGBClassifier(learning_rate = 0.3,
                             max_depth=5,
                             subsample = 1.0,
                             n_estimators=200)

xgb1_pca_model.fit(X_train_pca, y_train)

# Predictions on the train set
y_train_pred = xgb1_pca_model.predict(X_train_pca)

print("\nTraining set:")
model_metrics(y_train, y_train_pred)

# Prediction on the test set
y_test_pred = xgb1_pca_model.predict(X_test_pca)

print("\nValidation set:")
model_metrics(y_test, y_test_pred)

Training set:
+--------------------+-----------------+
| Metric             | Value           |
+====================+=================+
| Confusion matrix   | [[15255   145]  |
|                    |  [    8 15392]] |
+--------------------+-----------------+
| Accuracy           | 0.995           |
+--------------------+-----------------+
| Sensitivity/Recall | 0.999           |
+--------------------+-----------------+
| Specificity        | 0.991           |
+--------------------+-----------------+
| Precision          | 0.991           |
+--------------------+-----------------+
| F1-score           | 0.995           |
+--------------------+-----------------+

Validation set:
+--------------------+---------------+
| Metric             | Value         |
+====================+===============+
| Confusion matrix   | [[3587  276]  |
|                    |  [ 132  205]] |
+--------------------+---------------+
| Accuracy           | 0.903         |
+--------------------+---------------+
| Sensitivity/Recall | 0.608         |
+--------------------+---------------+
| Specificity        | 0.929         |
+--------------------+---------------+
| Precision          | 0.426         |
+--------------------+---------------+
| F1-score           | 0.501         |
+--------------------+---------------+

Model 5 - Random Forest Without PCA

While using PCA, we saw above how well Random Forest has performed. Therefore, we will use Random Forest to identify important features that cause a customer to churn.

rf = RandomForestClassifier(n_estimators=10, max_depth=4, max_features=5, random_state=100, oob_score=True)
rf.fit(X_train, y_train)
rf.oob_score_

0.8417532467532467

y_probs = rf.predict_proba(X_test)[:, 1]

# Calculate ROC curve
fpr, tpr, thresholds = roc_curve(y_test, y_probs)

# Calculate AUC (Area Under the Curve)
roc_auc = auc(fpr, tpr)

# Plot ROC curve
plt.figure(figsize=(6, 4))
plt.plot(fpr, tpr, color='darkorange', lw=2, label=f'ROC curve (AUC = {roc_auc:.2f})')
plt.plot([0, 1], [0, 1], color='navy', lw=2, linestyle='--', label='Random')
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('Receiver Operating Characteristic (ROC) Curve')
plt.legend(loc='lower right')
plt.show()

Hyperparameter tuning for Random Forest

# Tuning max_depth

params = {'max_depth': range(2, 40, 5)}
hpt(params,X_train,y_train,'roc_auc')

We will use 12 and 17 for max_depth.

# Tuning min_samples_leaf

params = {'min_samples_leaf': range(1, 100, 10)}
hpt(params,X_train,y_train,'roc_auc')

We will use values 10 and 20 for min_samples_leaf.

# Tuning min_samples_split
params = {'min_samples_split': range(10, 100, 10)}
hpt(params,X_train,y_train,'roc_auc')

We will use values between 20 and 30 for min_samples_split.

# Tuning n_estimators
params = {'n_estimators': range(50, 200, 50)}
hpt(params,X_train,y_train,'roc_auc')

We will use 100 estimators.

# Tuning max_features
params = {'max_features': range(10,50,10)}
hpt(params,X_train,y_train,'roc_auc',5)

We will use 15 max_features

folds = StratifiedKFold(n_splits = 3, shuffle = True, random_state = 100)
params = {
    'max_depth': [12,17],
    'min_samples_leaf': [10,20],
    'min_samples_split': [20,30],
    'n_estimators': [100], 
    'max_features': [15]
    }

# Create a based model
rfc = RandomForestClassifier(random_state=42, n_jobs=-1)

# Instantiate the grid search model
grid_search = GridSearchCV(estimator = rfc, 
                           param_grid = params, 
                           cv = folds,
                           n_jobs = -1,
                           verbose = 1, 
                           scoring = "roc_auc",
                           refit = True,
                           return_train_score=True)
# Fit the model
grid_search.fit(X_train, y_train)

Fitting 3 folds for each of 8 candidates, totalling 24 fits

GridSearchCV(cv=StratifiedKFold(n_splits=3, random_state=100, shuffle=True),
             estimator=RandomForestClassifier(n_jobs=-1, random_state=42),
             n_jobs=-1,
             param_grid={'max_depth': [12, 17], 'max_features': [15],
                         'min_samples_leaf': [10, 20],
                         'min_samples_split': [20, 30], 'n_estimators': [100]},
             return_train_score=True, scoring='roc_auc', verbose=1)

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

# print best hyperparameters
print("Best AUC: ", grid_search.best_score_)
print("Best hyperparameters: ", grid_search.best_params_)

Best AUC:  0.9804101070954854
Best hyperparameters:  {'max_depth': 17, 'max_features': 15, 'min_samples_leaf': 10, 'min_samples_split': 20, 'n_estimators': 100}

rfc1_model = RandomForestClassifier(bootstrap=True,
                             max_depth=17,
                             min_samples_leaf=10, 
                             min_samples_split=20,
                             max_features=15,
                             n_estimators=100)
rfc1_model.fit(X_train, y_train)

# Predictions on the train set
y_train_pred = rfc1_model.predict(X_train)
print("\nTraining set:")
model_metrics(y_train, y_train_pred)

# Prediction on the test set
y_test_pred = rfc1_model.predict(X_test)
print("\nValidation set:")
model_metrics(y_test, y_test_pred)

Training set:
+--------------------+-----------------+
| Metric             | Value           |
+====================+=================+
| Confusion matrix   | [[14743   657]  |
|                    |  [  332 15068]] |
+--------------------+-----------------+
| Accuracy           | 0.968           |
+--------------------+-----------------+
| Sensitivity/Recall | 0.978           |
+--------------------+-----------------+
| Specificity        | 0.957           |
+--------------------+-----------------+
| Precision          | 0.958           |
+--------------------+-----------------+
| F1-score           | 0.968           |
+--------------------+-----------------+

Validation set:
+--------------------+---------------+
| Metric             | Value         |
+====================+===============+
| Confusion matrix   | [[3603  260]  |
|                    |  [  91  246]] |
+--------------------+---------------+
| Accuracy           | 0.916         |
+--------------------+---------------+
| Sensitivity/Recall | 0.73          |
+--------------------+---------------+
| Specificity        | 0.933         |
+--------------------+---------------+
| Precision          | 0.486         |
+--------------------+---------------+
| F1-score           | 0.584         |
+--------------------+---------------+

y_probs = rfc1_model.predict_proba(X_test)[:, 1]

# Calculate ROC curve
fpr, tpr, thresholds = roc_curve(y_test, y_probs)

# Calculate AUC (Area Under the Curve)
roc_auc = auc(fpr, tpr)

# Plot ROC curve
plt.figure(figsize=(6, 4))
plt.plot(fpr, tpr, color='darkorange', lw=2, label=f'ROC curve (AUC = {roc_auc:.2f})')
plt.plot([0, 1], [0, 1], color='navy', lw=2, linestyle='--', label='Random')
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('Receiver Operating Characteristic (ROC) Curve')
plt.legend(loc='lower right')
plt.show()

rf_best = grid_search.best_estimator_
rf_best.feature_importances_

array([0.07429819, 0.00499139, 0.01862225, 0.06838385, 0.11619769,
       0.02126155, 0.03342574, 0.00193895, 0.0019986 , 0.00463145,
       0.00432121, 0.00036991, 0.00229602, 0.00417765, 0.        ,
       0.05996093, 0.11860825, 0.00411264, 0.00250917, 0.00581508,
       0.00146499, 0.00417055, 0.00222408, 0.00200574, 0.01133603,
       0.07389465, 0.04585511, 0.03200324, 0.03949407, 0.00857509,
       0.0088965 , 0.00314294, 0.00545635, 0.0018638 , 0.00045199,
       0.00569899, 0.00548926, 0.0283542 , 0.00220699, 0.00358296,
       0.00802657, 0.0042657 , 0.00466924, 0.00443595, 0.00562377,
       0.00389032, 0.00392108, 0.00586191, 0.00230565, 0.00496061,
       0.00531422, 0.00398655, 0.01061229, 0.00376367, 0.00729996,
       0.00471356, 0.00322928, 0.00217339, 0.00428024, 0.00559263,
       0.00485166, 0.00735617, 0.0043419 , 0.00416439, 0.0020631 ,
       0.00118487, 0.00722192, 0.0060248 , 0.00566229, 0.0036704 ,
       0.00516108, 0.00407848, 0.00407063, 0.00257502, 0.0045587 ])

imp_df = pd.DataFrame({
    "Varname": X_train.columns,
    "Imp": rf_best.feature_importances_
})
imp_df.sort_values(by="Imp", ascending=False)

	Varname	Imp
16	loc_ic_t2m_mou_8	0.118608
4	roam_og_mou_8	0.116198
0	arpu_8	0.074298
25	total_rech_amt_8	0.073895
3	roam_ic_mou_8	0.068384
15	loc_ic_t2t_mou_8	0.059961
26	max_rech_amt_8	0.045855
28	max_rech_data_8	0.039494
6	loc_og_t2m_mou_8	0.033426
27	last_day_rch_amt_8	0.032003
37	total_rech_amt_data_8	0.028354
5	loc_og_t2t_mou_8	0.021262
2	offnet_mou_8	0.018622
24	total_rech_num_8	0.011336
52	avg_roam_og_mou_6_7	0.010612
30	vol_3g_mb_8	0.008896
29	vol_2g_mb_8	0.008575
40	avg_std_og_t2t_mou_6_7	0.008027
61	avg_total_rech_num_6_7	0.007356
54	avg_loc_og_t2c_mou_6_7	0.007300
66	avg_loc_ic_t2m_mou_6_7	0.007222
67	avg_roam_ic_mou_6_7	0.006025
47	avg_total_rech_amt_6_7	0.005862
19	std_ic_t2m_mou_8	0.005815
35	aon	0.005699
68	avg_max_rech_amt_6_7	0.005662
44	avg_spl_og_mou_6_7	0.005624
59	avg_arpu_6_7	0.005593
36	3g_vbc_8	0.005489
32	sachet_2g_8	0.005456
50	avg_isd_ic_mou_6_7	0.005314
70	avg_max_rech_data_6_7	0.005161
1	onnet_mou_8	0.004991
49	avg_monthly_2g_6_7	0.004961
60	avg_last_day_rch_amt_6_7	0.004852
55	avg_std_og_t2m_mou_6_7	0.004714
42	avg_std_ic_t2t_mou_6_7	0.004669
9	std_og_t2t_mou_8	0.004631
74	avg_loc_ic_t2t_mou_6_7	0.004559
43	avg_3g_vbc_6_7	0.004436
62	avg_onnet_mou_6_7	0.004342
10	std_og_t2m_mou_8	0.004321
58	avg_loc_og_t2m_mou_6_7	0.004280
41	avg_total_rech_amt_data_6_7	0.004266
13	spl_og_mou_8	0.004178
21	spl_ic_mou_8	0.004171
63	avg_vol_3g_mb_6_7	0.004164
17	loc_ic_t2f_mou_8	0.004113
71	avg_offnet_mou_6_7	0.004078
72	avg_loc_og_t2t_mou_6_7	0.004071
51	avg_loc_og_t2f_mou_6_7	0.003987
46	avg_loc_ic_t2f_mou_6_7	0.003921
45	avg_ic_others_6_7	0.003890
53	avg_spl_ic_mou_6_7	0.003764
69	avg_vol_2g_mb_6_7	0.003670
39	avg_std_ic_t2m_mou_6_7	0.003583
56	avg_isd_og_mou_6_7	0.003229
31	monthly_2g_8	0.003143
73	avg_monthly_3g_6_7	0.002575
18	std_ic_t2t_mou_8	0.002509
48	avg_std_ic_t2f_mou_6_7	0.002306
12	isd_og_mou_8	0.002296
22	isd_ic_mou_8	0.002224
38	avg_og_others_6_7	0.002207
57	avg_sachet_2g_6_7	0.002173
64	avg_sachet_3g_6_7	0.002063
23	ic_others_8	0.002006
8	loc_og_t2c_mou_8	0.001999
7	loc_og_t2f_mou_8	0.001939
33	monthly_3g_8	0.001864
20	std_ic_t2f_mou_8	0.001465
65	avg_std_og_t2f_mou_6_7	0.001185
34	sachet_3g_8	0.000452
11	std_og_t2f_mou_8	0.000370
14	og_others_8	0.000000

Further finetuning hyperparameters for Random Forest

rfc2_model = RandomForestClassifier(bootstrap=True,
                             min_samples_split=15,
                             max_features='log2',
                             criterion='entropy',
                             class_weight='balanced',
                             n_estimators=150)
rfc2_model.fit(X_train, y_train)

# Predictions on the train set
y_train_pred = rfc2_model.predict(X_train)
print("\nTraining set:")
model_metrics(y_train, y_train_pred)

# Prediction on the test set
y_test_pred = rfc2_model.predict(X_test)
print("\nValidation set:")
model_metrics(y_test, y_test_pred)

Training set:
+--------------------+-----------------+
| Metric             | Value           |
+====================+=================+
| Confusion matrix   | [[15158   242]  |
|                    |  [   87 15313]] |
+--------------------+-----------------+
| Accuracy           | 0.989           |
+--------------------+-----------------+
| Sensitivity/Recall | 0.994           |
+--------------------+-----------------+
| Specificity        | 0.984           |
+--------------------+-----------------+
| Precision          | 0.984           |
+--------------------+-----------------+
| F1-score           | 0.989           |
+--------------------+-----------------+

Validation set:
+--------------------+---------------+
| Metric             | Value         |
+====================+===============+
| Confusion matrix   | [[3665  198]  |
|                    |  [ 102  235]] |
+--------------------+---------------+
| Accuracy           | 0.929         |
+--------------------+---------------+
| Sensitivity/Recall | 0.697         |
+--------------------+---------------+
| Specificity        | 0.949         |
+--------------------+---------------+
| Precision          | 0.543         |
+--------------------+---------------+
| F1-score           | 0.61          |
+--------------------+---------------+

y_probs = rfc2_model.predict_proba(X_test)[:, 1]

# Calculate ROC curve
fpr, tpr, thresholds = roc_curve(y_test, y_probs)

# Calculate AUC (Area Under the Curve)
roc_auc = auc(fpr, tpr)

# Plot ROC curve
plt.figure(figsize=(6, 4))
plt.plot(fpr, tpr, color='darkorange', lw=2, label=f'ROC curve (AUC = {roc_auc:.2f})')
plt.plot([0, 1], [0, 1], color='navy', lw=2, linestyle='--', label='Random')
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('Receiver Operating Characteristic (ROC) Curve')
plt.legend(loc='lower right')
plt.show()

The above ROC curve follows closer to the left-hand border and then the top border of the ROC space, therefore this is a more accurate test.

# More finetuning

rfc3_model = RandomForestClassifier(bootstrap=True,
                             min_samples_split=15,
                             max_features='log2',
                             criterion='entropy',
                             class_weight='balanced',
                             n_estimators=150)
rfc3_model.fit(X_train, y_train)

# Predictions on the train set
y_train_pred = rfc3_model.predict(X_train)
print("\nTraining set:")
model_metrics(y_train, y_train_pred)

# Prediction on the test set
y_test_pred = rfc3_model.predict(X_test)
print("\nValidation set:")
model_metrics(y_test, y_test_pred)

Training set:
+--------------------+-----------------+
| Metric             | Value           |
+====================+=================+
| Confusion matrix   | [[15160   240]  |
|                    |  [   88 15312]] |
+--------------------+-----------------+
| Accuracy           | 0.989           |
+--------------------+-----------------+
| Sensitivity/Recall | 0.994           |
+--------------------+-----------------+
| Specificity        | 0.984           |
+--------------------+-----------------+
| Precision          | 0.985           |
+--------------------+-----------------+
| F1-score           | 0.989           |
+--------------------+-----------------+

Validation set:
+--------------------+---------------+
| Metric             | Value         |
+====================+===============+
| Confusion matrix   | [[3654  209]  |
|                    |  [ 104  233]] |
+--------------------+---------------+
| Accuracy           | 0.925         |
+--------------------+---------------+
| Sensitivity/Recall | 0.691         |
+--------------------+---------------+
| Specificity        | 0.946         |
+--------------------+---------------+
| Precision          | 0.527         |
+--------------------+---------------+
| F1-score           | 0.598         |
+--------------------+---------------+

# More finetuning

rfc4_model = RandomForestClassifier(
                             #min_samples_split=15,
                             max_features='log2',
                             #criterion='entropy',
                             class_weight='balanced',
                             n_estimators=500)
rfc4_model.fit(X_train, y_train)

# Predictions on the train set
y_train_pred = rfc4_model.predict(X_train)
print("\nTraining set:")
model_metrics(y_train, y_train_pred)

# Prediction on the test set
y_test_pred = rfc4_model.predict(X_test)
print("\nValidation set:")
model_metrics(y_test, y_test_pred)

Training set:
+--------------------+-----------------+
| Metric             | Value           |
+====================+=================+
| Confusion matrix   | [[15400     0]  |
|                    |  [    0 15400]] |
+--------------------+-----------------+
| Accuracy           | 1.0             |
+--------------------+-----------------+
| Sensitivity/Recall | 1.0             |
+--------------------+-----------------+
| Specificity        | 1.0             |
+--------------------+-----------------+
| Precision          | 1.0             |
+--------------------+-----------------+
| F1-score           | 1.0             |
+--------------------+-----------------+

Validation set:
+--------------------+---------------+
| Metric             | Value         |
+====================+===============+
| Confusion matrix   | [[3698  165]  |
|                    |  [ 109  228]] |
+--------------------+---------------+
| Accuracy           | 0.935         |
+--------------------+---------------+
| Sensitivity/Recall | 0.677         |
+--------------------+---------------+
| Specificity        | 0.957         |
+--------------------+---------------+
| Precision          | 0.58          |
+--------------------+---------------+
| F1-score           | 0.625         |
+--------------------+---------------+

STEP 5: Model Recommendation

1. Random Forest and SVM both perform well on the given dataset. Models with PCA and without PCA have been executed.
2. Random Forest models are not just accurate but also perform better than SVM on other metrics like sensitivity and specificty. Therefore, we will choose rfc4_model for predictions on the test data. It will be used to predict whether a high-value customer will churn or not, in near future (i.e. churn phase). By knowing this, the company can take action steps such as providing special plans, discounts on recharge etc.
3. The model rfc1_model is being used to identify the top predictors of customer churn. It will be used to identify important variables that are strong predictors of churn. These variables may also indicate why customers choose to switch to other networks. The top predictors are displayed below in the next section.
4. log1_pca_model - This is a logistic regression model. Although this model has lower accuracy compared to other models, the sensitivity of this model is more than 80% which is the highest among all models. We would recommend this model as it identifies customers who'll definitely churn with more accuracy as compared to the ones who'll not churn. This is given by the metric sensitivity. Logistic regression models are simple and consume lesser computational resource.

Top Predictors of Churn

• It can be clearly seen that attributes pertaining to month 8 (August) impact the probability of churning heavily. This is something we observed even during EDA while plotting the scatter plots.
• Another important predictor of churn is Age on Network.
• The top predictors of churn as given by the model are as below.

imp_df.sort_values(by="Imp", ascending=False)

	Varname	Imp
16	loc_ic_t2m_mou_8	0.118608
4	roam_og_mou_8	0.116198
0	arpu_8	0.074298
25	total_rech_amt_8	0.073895
3	roam_ic_mou_8	0.068384
15	loc_ic_t2t_mou_8	0.059961
26	max_rech_amt_8	0.045855
28	max_rech_data_8	0.039494
6	loc_og_t2m_mou_8	0.033426
27	last_day_rch_amt_8	0.032003
37	total_rech_amt_data_8	0.028354
5	loc_og_t2t_mou_8	0.021262
2	offnet_mou_8	0.018622
24	total_rech_num_8	0.011336
52	avg_roam_og_mou_6_7	0.010612
30	vol_3g_mb_8	0.008896
29	vol_2g_mb_8	0.008575
40	avg_std_og_t2t_mou_6_7	0.008027
61	avg_total_rech_num_6_7	0.007356
54	avg_loc_og_t2c_mou_6_7	0.007300
66	avg_loc_ic_t2m_mou_6_7	0.007222
67	avg_roam_ic_mou_6_7	0.006025
47	avg_total_rech_amt_6_7	0.005862
19	std_ic_t2m_mou_8	0.005815
35	aon	0.005699
68	avg_max_rech_amt_6_7	0.005662
44	avg_spl_og_mou_6_7	0.005624
59	avg_arpu_6_7	0.005593
36	3g_vbc_8	0.005489
32	sachet_2g_8	0.005456
50	avg_isd_ic_mou_6_7	0.005314
70	avg_max_rech_data_6_7	0.005161
1	onnet_mou_8	0.004991
49	avg_monthly_2g_6_7	0.004961
60	avg_last_day_rch_amt_6_7	0.004852
55	avg_std_og_t2m_mou_6_7	0.004714
42	avg_std_ic_t2t_mou_6_7	0.004669
9	std_og_t2t_mou_8	0.004631
74	avg_loc_ic_t2t_mou_6_7	0.004559
43	avg_3g_vbc_6_7	0.004436
62	avg_onnet_mou_6_7	0.004342
10	std_og_t2m_mou_8	0.004321
58	avg_loc_og_t2m_mou_6_7	0.004280
41	avg_total_rech_amt_data_6_7	0.004266
13	spl_og_mou_8	0.004178
21	spl_ic_mou_8	0.004171
63	avg_vol_3g_mb_6_7	0.004164
17	loc_ic_t2f_mou_8	0.004113
71	avg_offnet_mou_6_7	0.004078
72	avg_loc_og_t2t_mou_6_7	0.004071
51	avg_loc_og_t2f_mou_6_7	0.003987
46	avg_loc_ic_t2f_mou_6_7	0.003921
45	avg_ic_others_6_7	0.003890
53	avg_spl_ic_mou_6_7	0.003764
69	avg_vol_2g_mb_6_7	0.003670
39	avg_std_ic_t2m_mou_6_7	0.003583
56	avg_isd_og_mou_6_7	0.003229
31	monthly_2g_8	0.003143
73	avg_monthly_3g_6_7	0.002575
18	std_ic_t2t_mou_8	0.002509
48	avg_std_ic_t2f_mou_6_7	0.002306
12	isd_og_mou_8	0.002296
22	isd_ic_mou_8	0.002224
38	avg_og_others_6_7	0.002207
57	avg_sachet_2g_6_7	0.002173
64	avg_sachet_3g_6_7	0.002063
23	ic_others_8	0.002006
8	loc_og_t2c_mou_8	0.001999
7	loc_og_t2f_mou_8	0.001939
33	monthly_3g_8	0.001864
20	std_ic_t2f_mou_8	0.001465
65	avg_std_og_t2f_mou_6_7	0.001185
34	sachet_3g_8	0.000452
11	std_og_t2f_mou_8	0.000370
14	og_others_8	0.000000

Business Recommendations

From EDA, we observe the following:

• When people have high total charges, they tend to churn less. This means, when they have greater liability towards the operator, they do not churn.
  • Recommendation: Bring in long term packs like half yearly or yearly ones at attractive prices - Customers purchasing these packs will tend to churn less.
• Likewise, when people have stayed longer with the operator (>2 years), they churn less.
  • Recommendation: For new customers, provide attractive offers and schemes that will make them stay longer with the operator.
• Call related attributes have a greater impact on churn than data related packs.
  • Recommendation: Invest into bring attractive call-related packs. Whenever customers buy smaller call related packs, the operator must be alerted for a potential churn.
• Unusual dip in purchase of packs or usage of services before churning can be seen. We see people who have suddenly reduced usage in the 8th month in comparison to 6th and 7th months have churned more.
  • Recommendation: Whenever a sudden dip in usage is noticed, the operator must be alerted for a potential churn.

STEP 6: Kaggle Submission

kaggle = pd.read_csv("test.csv")
kaggle.head()

	id	circle_id	loc_og_t2o_mou	std_og_t2o_mou	loc_ic_t2o_mou	last_date_of_month_6	last_date_of_month_7	last_date_of_month_8	arpu_6	arpu_7	arpu_8	onnet_mou_6	onnet_mou_7	onnet_mou_8	offnet_mou_6	offnet_mou_7	offnet_mou_8	roam_ic_mou_6	roam_ic_mou_7	roam_ic_mou_8	roam_og_mou_6	roam_og_mou_7	roam_og_mou_8	loc_og_t2t_mou_6	loc_og_t2t_mou_7	loc_og_t2t_mou_8	loc_og_t2m_mou_6	loc_og_t2m_mou_7	loc_og_t2m_mou_8	loc_og_t2f_mou_6	loc_og_t2f_mou_7	loc_og_t2f_mou_8	loc_og_t2c_mou_6	loc_og_t2c_mou_7	loc_og_t2c_mou_8	loc_og_mou_6	loc_og_mou_7	loc_og_mou_8	std_og_t2t_mou_6	std_og_t2t_mou_7	std_og_t2t_mou_8	std_og_t2m_mou_6	std_og_t2m_mou_7	std_og_t2m_mou_8	std_og_t2f_mou_6	std_og_t2f_mou_7	std_og_t2f_mou_8	std_og_t2c_mou_6	std_og_t2c_mou_7	std_og_t2c_mou_8	std_og_mou_6	std_og_mou_7	std_og_mou_8	isd_og_mou_6	isd_og_mou_7	isd_og_mou_8	spl_og_mou_6	spl_og_mou_7	spl_og_mou_8	og_others_6	og_others_7	og_others_8	total_og_mou_6	total_og_mou_7	total_og_mou_8	loc_ic_t2t_mou_6	loc_ic_t2t_mou_7	loc_ic_t2t_mou_8	loc_ic_t2m_mou_6	loc_ic_t2m_mou_7	loc_ic_t2m_mou_8	loc_ic_t2f_mou_6	loc_ic_t2f_mou_7	loc_ic_t2f_mou_8	loc_ic_mou_6	loc_ic_mou_7	loc_ic_mou_8	std_ic_t2t_mou_6	std_ic_t2t_mou_7	std_ic_t2t_mou_8	std_ic_t2m_mou_6	std_ic_t2m_mou_7	std_ic_t2m_mou_8	std_ic_t2f_mou_6	std_ic_t2f_mou_7	std_ic_t2f_mou_8	std_ic_t2o_mou_6	std_ic_t2o_mou_7	std_ic_t2o_mou_8	std_ic_mou_6	std_ic_mou_7	std_ic_mou_8	total_ic_mou_6	total_ic_mou_7	total_ic_mou_8	spl_ic_mou_6	spl_ic_mou_7	spl_ic_mou_8	isd_ic_mou_6	isd_ic_mou_7	isd_ic_mou_8	ic_others_6	ic_others_7	ic_others_8	total_rech_num_6	total_rech_num_7	total_rech_num_8	total_rech_amt_6	total_rech_amt_7	total_rech_amt_8	max_rech_amt_6	max_rech_amt_7	max_rech_amt_8	date_of_last_rech_6	date_of_last_rech_7	date_of_last_rech_8	last_day_rch_amt_6	last_day_rch_amt_7	last_day_rch_amt_8	date_of_last_rech_data_6	date_of_last_rech_data_7	date_of_last_rech_data_8	total_rech_data_6	total_rech_data_7	total_rech_data_8	max_rech_data_6	max_rech_data_7	max_rech_data_8	count_rech_2g_6	count_rech_2g_7	count_rech_2g_8	count_rech_3g_6	count_rech_3g_7	count_rech_3g_8	av_rech_amt_data_6	av_rech_amt_data_7	av_rech_amt_data_8	vol_2g_mb_6	vol_2g_mb_7	vol_2g_mb_8	vol_3g_mb_6	vol_3g_mb_7	vol_3g_mb_8	arpu_3g_6	arpu_3g_7	arpu_3g_8	arpu_2g_6	arpu_2g_7	arpu_2g_8	night_pck_user_6	night_pck_user_7	night_pck_user_8	monthly_2g_6	monthly_2g_7	monthly_2g_8	sachet_2g_6	sachet_2g_7	sachet_2g_8	monthly_3g_6	monthly_3g_7	monthly_3g_8	sachet_3g_6	sachet_3g_7	sachet_3g_8	fb_user_6	fb_user_7	fb_user_8	aon	aug_vbc_3g	jul_vbc_3g	jun_vbc_3g
0	69999	109	0.0	0.0	0.0	6/30/2014	7/31/2014	8/31/2014	91.882	65.330	64.445	31.78	20.23	23.11	60.16	32.16	34.83	0.00	0.00	0.00	0.00	0.00	0.00	24.88	20.23	21.06	18.13	10.89	8.36	0.00	13.58	0.00	0.0	0.00	0.03	43.01	44.71	29.43	6.90	0.00	2.05	42.03	7.68	26.43	0.0	0.0	0.0	0.0	0.0	0.0	48.93	7.68	28.48	0.0	0.0	0.0	0.00	0.00	0.03	0.0	0.0	0.0	91.94	52.39	57.94	30.33	37.56	21.98	10.21	4.59	9.53	0.26	0.00	0.00	40.81	42.16	31.51	0.00	0.00	0.00	0.36	1.04	4.34	0.0	0.0	0.00	0.0	0.0	0.0	0.36	1.04	4.34	41.73	43.56	36.26	0.54	0.34	0.39	0.00	0.00	0.00	0.0	0.0	0.00	5	5	4	103	90	60	50	30	30	6/21/2014	7/26/2014	8/24/2014	30	30	0	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	0.0	0.0	0.0	0.0	0.0	0.0	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	0	0	0	0	0	0	0	0	0	0	0	0	NaN	NaN	NaN	1692	0.00	0.00	0.00
1	70000	109	0.0	0.0	0.0	6/30/2014	7/31/2014	8/31/2014	414.168	515.568	360.868	75.51	41.21	19.84	474.34	621.84	394.94	0.00	0.00	0.00	0.00	0.00	0.00	75.51	41.21	19.84	473.61	598.08	377.26	0.73	0.00	0.00	0.0	0.00	0.00	549.86	639.29	397.11	0.00	0.00	0.00	0.00	23.76	17.68	0.0	0.0	0.0	0.0	0.0	0.0	0.00	23.76	17.68	0.0	0.0	0.8	0.00	0.00	0.00	0.0	0.0	0.0	549.86	663.06	415.59	19.99	26.95	2.61	160.19	122.29	184.81	1.49	0.00	0.00	181.69	149.24	187.43	0.00	0.00	0.00	0.00	12.51	0.00	0.0	0.0	0.00	0.0	0.0	0.0	0.00	12.51	0.00	296.33	339.64	281.66	0.00	0.00	0.00	114.63	177.88	94.23	0.0	0.0	0.00	5	4	5	500	500	500	250	250	250	6/19/2014	7/16/2014	8/24/2014	250	0	0	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	0.0	0.0	0.0	0.0	0.0	0.0	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	0	0	0	0	0	0	0	0	0	0	0	0	NaN	NaN	NaN	2533	0.00	0.00	0.00
2	70001	109	0.0	0.0	0.0	6/30/2014	7/31/2014	8/31/2014	329.844	434.884	746.239	7.54	7.86	8.40	16.98	45.81	45.04	22.81	103.38	26.08	24.53	53.68	54.44	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.0	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.0	0.0	0.0	0.0	0.0	0.0	0.00	0.00	0.00	0.0	0.0	0.0	0.00	0.00	0.00	0.0	0.0	0.0	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.0	0.0	0.00	0.0	0.0	0.0	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.0	0.0	0.00	6	9	5	500	1000	1000	300	500	500	6/29/2014	7/27/2014	8/28/2014	0	0	0	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	0.0	0.0	0.0	0.0	0.0	0.0	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	0	0	0	0	0	0	0	0	0	0	0	0	NaN	NaN	NaN	277	525.61	758.41	241.84
3	70002	109	0.0	0.0	0.0	6/30/2014	7/31/2014	8/31/2014	43.550	171.390	24.400	5.31	2.16	0.00	40.04	205.01	24.01	0.00	0.00	0.00	0.00	0.00	0.00	5.31	0.00	0.00	2.94	98.61	20.51	0.00	0.00	2.35	0.0	6.18	0.00	8.26	98.61	22.86	0.00	2.16	0.00	37.09	94.36	0.00	0.0	0.0	0.0	0.0	0.0	0.0	37.09	96.53	0.00	0.0	0.0	0.0	0.00	12.03	1.15	0.0	0.0	0.0	45.36	207.18	24.01	58.11	54.64	23.04	487.94	449.83	506.94	0.00	0.38	1.64	546.06	504.86	531.64	0.00	4.26	0.00	9.63	11.88	8.83	0.0	0.0	0.00	0.0	0.0	0.0	9.63	16.14	8.83	555.69	522.44	549.13	0.00	0.00	0.00	0.00	1.43	8.65	0.0	0.0	0.00	3	5	2	110	260	0	110	150	0	6/25/2014	7/30/2014	8/24/2014	110	150	0	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	0.0	0.0	0.0	0.0	0.0	0.0	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	0	0	0	0	0	0	0	0	0	0	0	0	NaN	NaN	NaN	1244	0.00	0.00	0.00
4	70003	109	0.0	0.0	0.0	6/30/2014	7/31/2014	8/31/2014	306.854	406.289	413.329	450.93	609.03	700.68	60.94	23.84	74.16	0.00	0.00	0.00	0.00	0.00	0.00	0.45	0.78	14.56	2.39	2.66	10.94	0.00	0.00	0.00	0.0	0.00	0.00	2.84	3.44	25.51	450.48	608.24	686.11	58.54	21.18	63.18	0.0	0.0	0.0	0.0	0.0	0.0	509.03	629.43	749.29	0.0	0.0	0.0	0.71	5.39	4.96	2.2	0.0	0.0	514.79	638.28	779.78	0.00	0.36	9.91	10.13	9.23	7.69	0.00	0.00	0.00	10.13	9.59	17.61	29.71	92.36	107.39	13.88	13.96	32.46	0.0	0.0	1.61	0.0	0.0	0.0	43.59	106.33	141.48	53.73	115.93	159.26	0.00	0.00	0.00	0.00	0.00	0.00	0.0	0.0	0.16	11	7	8	356	490	546	90	130	130	6/29/2014	7/29/2014	8/30/2014	50	130	130	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	0.0	0.0	0.0	0.0	0.0	0.0	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	0	0	0	0	0	0	0	0	0	0	0	0	NaN	NaN	NaN	462	0.00	0.00	0.00

kaggle.shape

(30000, 171)

# Column cleanup

# Rename columns
kaggle = kaggle.rename(columns = {'aug_vbc_3g':'3g_vbc_8','jul_vbc_3g':'3g_vbc_7','jun_vbc_3g':'3g_vbc_6'})

# Impute 0 for data recharge columns
kaggle['total_rech_data_6'] = kaggle['total_rech_data_6'].replace(np.NaN,0.0)
kaggle['total_rech_data_7'] = kaggle['total_rech_data_7'].replace(np.NaN,0.0)
kaggle['total_rech_data_8'] = kaggle['total_rech_data_8'].replace(np.NaN,0.0)
kaggle['av_rech_amt_data_6'] = kaggle['av_rech_amt_data_6'].replace(np.NaN,0.0)
kaggle['av_rech_amt_data_7'] = kaggle['av_rech_amt_data_7'].replace(np.NaN,0.0)
kaggle['av_rech_amt_data_8'] = kaggle['av_rech_amt_data_8'].replace(np.NaN,0.0)

# Create new column: total recharge amount for data: total_rech_amt_data 
kaggle['total_rech_amt_data_6'] = kaggle.av_rech_amt_data_6 * kaggle.total_rech_data_6
kaggle['total_rech_amt_data_7'] = kaggle.av_rech_amt_data_7 * kaggle.total_rech_data_7
kaggle['total_rech_amt_data_8'] = kaggle.av_rech_amt_data_8 * kaggle.total_rech_data_8

# Impute 0 for max data recharge columns
kaggle['max_rech_data_6'] = kaggle['max_rech_data_6'].replace(np.NaN,0.0)
kaggle['max_rech_data_7'] = kaggle['max_rech_data_7'].replace(np.NaN,0.0)
kaggle['max_rech_data_8'] = kaggle['max_rech_data_8'].replace(np.NaN,0.0)

# Impute Minutes of Usage column
kaggle[col_mou] = kaggle[col_mou].replace(np.NaN,0.0)

# Impute incoming/outgoing columns
kaggle[col_og_ic] = kaggle[col_og_ic].replace(np.NaN,0.0)

# Average out months 6 and 7
for idx, col in enumerate(col_6_7_common):
    avg_col_6_7 = "avg_"+col+"_6_7"
    col_6 = col+"_6"
    col_7 = col+"_7"
    kaggle[avg_col_6_7] = (kaggle[col_6]  + kaggle[col_7])/ 2

kaggle_pred = kaggle[final_column_list].copy()
kaggle_pred.shape

(30000, 75)

# Outlier Treatment
pct_change_99_1 = kaggle_pred.quantile(np.arange(0.9,1.01,0.01),axis=0, numeric_only=True).pct_change().mul(100).iloc[-1]
outlier_condition = pct_change_99_1 > 100
columns_with_outliers = pct_change_99_1[outlier_condition].index.values
for col in columns_with_outliers : 
    outlier_threshold = kaggle_pred[col].quantile(0.99)
    condition = kaggle_pred[col] > outlier_threshold
    kaggle_pred.loc[condition, col] = outlier_threshold

# Scale
kaggle_pred[final_column_list] = scaler.transform(kaggle_pred[final_column_list])
kaggle_pred.head()

	arpu_8	onnet_mou_8	offnet_mou_8	roam_ic_mou_8	roam_og_mou_8	loc_og_t2t_mou_8	loc_og_t2m_mou_8	loc_og_t2f_mou_8	loc_og_t2c_mou_8	std_og_t2t_mou_8	std_og_t2m_mou_8	std_og_t2f_mou_8	isd_og_mou_8	spl_og_mou_8	og_others_8	loc_ic_t2t_mou_8	loc_ic_t2m_mou_8	loc_ic_t2f_mou_8	std_ic_t2t_mou_8	std_ic_t2m_mou_8	std_ic_t2f_mou_8	spl_ic_mou_8	isd_ic_mou_8	ic_others_8	total_rech_num_8	total_rech_amt_8	max_rech_amt_8	last_day_rch_amt_8	max_rech_data_8	vol_2g_mb_8	vol_3g_mb_8	monthly_2g_8	sachet_2g_8	monthly_3g_8	sachet_3g_8	aon	3g_vbc_8	total_rech_amt_data_8	avg_og_others_6_7	avg_std_ic_t2m_mou_6_7	avg_std_og_t2t_mou_6_7	avg_total_rech_amt_data_6_7	avg_std_ic_t2t_mou_6_7	avg_3g_vbc_6_7	avg_spl_og_mou_6_7	avg_ic_others_6_7	avg_loc_ic_t2f_mou_6_7	avg_total_rech_amt_6_7	avg_std_ic_t2f_mou_6_7	avg_monthly_2g_6_7	avg_isd_ic_mou_6_7	avg_loc_og_t2f_mou_6_7	avg_roam_og_mou_6_7	avg_spl_ic_mou_6_7	avg_loc_og_t2c_mou_6_7	avg_std_og_t2m_mou_6_7	avg_isd_og_mou_6_7	avg_sachet_2g_6_7	avg_loc_og_t2m_mou_6_7	avg_arpu_6_7	avg_last_day_rch_amt_6_7	avg_total_rech_num_6_7	avg_onnet_mou_6_7	avg_vol_3g_mb_6_7	avg_sachet_3g_6_7	avg_std_og_t2f_mou_6_7	avg_loc_ic_t2m_mou_6_7	avg_roam_ic_mou_6_7	avg_max_rech_amt_6_7	avg_vol_2g_mb_6_7	avg_max_rech_data_6_7	avg_offnet_mou_6_7	avg_loc_og_t2t_mou_6_7	avg_monthly_3g_6_7	avg_loc_ic_t2t_mou_6_7
0	-0.789460	-0.418184	-0.541440	-0.309217	-0.345146	-0.209192	-0.502761	-0.327195	-0.297952	-0.360552	-0.307597	-0.187171	-0.152370	-0.418148	0.0	-0.179451	-0.530555	-0.357595	-0.315194	-0.268584	-0.224676	5.169981	-0.229021	-0.216127	-0.443756	-0.772632	-0.629697	-0.636077	-0.530001	-0.307769	-0.356159	-0.277974	-0.337643	-0.281795	-0.190088	0.720061	-0.351276	-0.344829	-0.430833	-0.498593	-0.517514	-0.515336	-0.450548	-0.430986	-0.609061	-0.312114	-0.422147	-1.471579	-0.293179	-0.41431	-0.306304	0.204236	-0.366103	5.128793	-0.450335	-0.496410	-0.212023	-0.492124	-0.647763	-1.479153	-0.688108	-0.902716	-0.604792	-0.470776	-0.326582	-0.245369	-0.742431	-0.344064	-0.941955	-0.455492	-0.743385	-0.766474	-0.305738	-0.424651	-0.183716
1	-0.013426	-0.427387	0.434662	-0.309217	-0.345146	-0.220124	1.733102	-0.327195	-0.306055	-0.367277	-0.337788	-0.187171	0.114281	-0.421099	0.0	-0.429804	0.661045	-0.357595	-0.315194	-0.367630	-0.224676	-0.204051	3.488968	-0.216127	-0.311227	0.190799	0.926188	-0.636077	-0.530001	-0.307769	-0.356159	-0.277974	-0.337643	-0.281795	-0.190088	1.671295	-0.351276	-0.344829	-0.430833	-0.382935	-0.526533	-0.515336	-0.450548	-0.430986	-0.609061	-0.312114	-0.398591	-0.404664	-0.293179	-0.41431	4.989777	-0.364628	-0.366103	-0.481483	-0.450335	-0.530292	-0.212023	-0.492124	2.345319	-0.279376	0.368238	-0.967979	-0.528701	-0.470776	-0.326582	-0.245369	0.162478	-0.344064	0.707954	-0.455492	-0.743385	0.386454	-0.025705	-0.424651	-0.309296
2	0.995474	-0.459583	-0.513765	0.371496	0.545028	-0.397901	-0.553430	-0.327195	-0.306055	-0.367277	-0.398793	-0.187171	-0.152370	-0.421099	0.0	-0.463537	-0.595343	-0.357595	-0.315194	-0.367630	-0.224676	-0.204051	-0.229021	-0.216127	-0.311227	1.285606	2.694240	-0.636077	-0.530001	-0.307769	-0.356159	-0.277974	-0.337643	-0.281795	-0.190088	-0.880409	1.513000	-0.344829	-0.430833	-0.513168	-0.526533	-0.515336	-0.450548	1.241259	-0.609061	-0.312114	-0.427126	0.256373	-0.293179	-0.41431	-0.306304	-0.396945	0.171186	-0.481483	-0.450335	-0.561314	-0.212023	-0.492124	-0.731067	-0.535643	-1.021691	-0.576404	-0.647840	-0.470776	-0.326582	-0.245369	-0.792464	1.196480	1.886460	-0.455492	-0.743385	-0.800389	-0.482141	-0.424651	-0.590667
3	-0.894298	-0.483224	-0.570768	-0.309217	-0.345146	-0.397901	-0.429121	-0.103626	-0.306055	-0.367277	-0.398793	-0.187171	-0.152370	-0.307992	0.0	-0.165751	2.850972	-0.284834	-0.315194	-0.166115	-0.224676	-0.204051	0.112278	-0.216127	-0.708813	-0.904008	-0.841863	-0.636077	-0.530001	-0.307769	-0.356159	-0.277974	-0.337643	-0.281795	-0.190088	0.213340	-0.351276	-0.344829	-0.430833	-0.289242	-0.523710	-0.515336	-0.371494	-0.430986	0.004377	-0.312114	-0.419849	-1.237572	-0.293179	-0.41431	-0.280412	-0.396945	-0.366103	-0.481483	0.499941	-0.389686	-0.212023	-0.492124	-0.439558	-1.389498	0.423835	-1.033241	-0.657165	-0.470776	-0.326582	-0.245369	2.377728	-0.344064	-0.234851	-0.455492	-0.743385	-0.591064	-0.461377	-0.424651	0.085186
4	0.123917	1.488731	-0.434834	-0.309217	-0.345146	-0.267435	-0.487123	-0.327195	-0.306055	1.883568	-0.180792	-0.187171	-0.152370	0.066736	0.0	-0.335453	-0.543064	-0.357595	4.106433	0.373158	0.092549	-0.204051	-0.229021	-0.147439	0.086359	0.291521	0.077524	0.711676	-0.530001	-0.307769	-0.356159	-0.277974	-0.337643	-0.281795	-0.190088	-0.671160	-0.351276	-0.344829	0.982048	-0.223345	0.857184	-0.515336	1.814744	-0.430986	-0.298008	-0.312114	-0.427126	-0.608264	-0.293179	-0.41431	-0.306304	-0.396945	-0.366103	-0.481483	-0.450335	-0.457227	-0.212023	-0.492124	-0.716571	-0.615758	-0.020942	-0.380617	0.580430	-0.470776	-0.326582	-0.245369	-0.727016	-0.344064	-0.391985	-0.455492	-0.743385	-0.775134	-0.477332	-0.424651	-0.588509

# Fetch ids for kaggle upload
kaggle_id = kaggle['id']

# Model building with PCA
kaggle_pred_pca = pca_final.transform(kaggle_pred)

#1 Logistic Regression

log1_pred_kaggle = log1_pca_model.predict(kaggle_pred_pca)
result = pd.DataFrame({"id":kaggle_id,"churn_probability":log1_pred_kaggle})
result.to_csv('Submission_log.csv', header=True,index=False)

#2 SVM

svm1_pred_kaggle = svm1_pca_model.predict(kaggle_pred_pca)
result = pd.DataFrame({"id":kaggle_id,"churn_probability":svm1_pred_kaggle})
result.to_csv('Submission_svm.csv', header=True,index=False)

#3 Random Forest

#4 Random Forest
rf_pred_kaggle = rf3_pca_model.predict(kaggle_pred_pca)
result = pd.DataFrame({"id":kaggle_id,"churn_probability":rf_pred_kaggle})
result.to_csv('Submission_rf.csv', header=True,index=False)

#4 XGBoost

#4 XGB
xgb_pred_kaggle = xgb1_pca_model.predict(kaggle_pred_pca)
result = pd.DataFrame({"id":kaggle_id,"churn_probability":xgb_pred_kaggle})
result.to_csv('Submission_xgb.csv', header=True,index=False)

#5 Random Forest Without PCA

rfc1_pred_kaggle = rfc1_model.predict(kaggle_pred)
result = pd.DataFrame({"id":kaggle_id,"churn_probability":rfc1_pred_kaggle})
result.to_csv('Submission_rf_1.csv', header=True,index=False)

rfc2_pred_kaggle = rfc2_model.predict(kaggle_pred)
result = pd.DataFrame({"id":kaggle_id,"churn_probability":rfc2_pred_kaggle})
result.to_csv('Submission_rf_2.csv', header=True,index=False)

rfc3_pred_kaggle = rfc3_model.predict(kaggle_pred)
result = pd.DataFrame({"id":kaggle_id,"churn_probability":rfc3_pred_kaggle})
result.to_csv('Submission_rf_4.csv', header=True,index=False)

rfc4_pred_kaggle = rfc4_model.predict(kaggle_pred)
result = pd.DataFrame({"id":kaggle_id,"churn_probability":rfc4_pred_kaggle})
result.to_csv('Submission_rf_5.csv', header=True,index=False)